Download the Book (PDF): Introduction In the twelve months that ended in the late summer of 2023, more than 110,000 people in the United States died of drug overdoses. No previous year in the nation's history came close. To put the number in clinical terms, it was more than the deaths from motor vehicle crashes and firearms combined, and it fell disproportionately on people in the middle of their working lives. Then, with a speed almost nobody predicted, the curve broke. By the Centers for Disease Control and Prevention's provisional count, overdose deaths fell to roughly 81,000 in 2024 and to just under 70,000 in 2025, the third consecutive annual decline. Opioid-involved deaths fell faster still, by close to a fifth in a single year. Seventy thousand deaths is not a victory. It is roughly where the country stood in 2017, when the epidemic was already being called the worst public-health catastrophe of a generation. But the decline matters, because it disposes of a fatalism that had settled over the field. For most of the 2010s, clinicians and policymakers spoke about overdose as though it were weather: a force that arrived from outside, shifted with the drug supply, and could be endured but not governed. The recent fall, uneven as it is across states and populations, suggests otherwise. Something in the system's response has begun to work, and it is worth understanding precisely what. This book makes a single argument. The opioid crisis was produced in large part by medicine, and it is being unmade, slowly and incompletely, by medicine that has learned to treat opioid use disorder as what it is: a chronic, relapsing, frequently fatal illness for which effective treatment exists. The tragedy of the past three decades is not that the epidemic was mysterious. It is that for most of that period the tools that reduce death were known, cheap, and available, yet were kept scarce by regulation, stigma, and institutional habit. The policy task of the coming decade is less about inventing new interventions than about removing the barriers that still stand between people at risk of dying and the treatments that keep them alive. Why "iatrogenic" is the right word The word iatrogenic means harm caused by the healer. It is sometimes used loosely in this field, and it can be misread as a claim that every person who developed an opioid use disorder was started on the path by a prescription. That is not what the evidence shows, and this book does not argue it. Many people who died of overdose in the fentanyl era never had a legitimate opioid prescription in their lives. The claim is structural rather than individual. Beginning in the mid-1990s, the American medical system, encouraged by pharmaceutical marketing and by a genuine, well-intentioned reform movement in pain care, dramatically expanded the volume of potent opioids flowing into households. Prescriptions peaked in 2012 at a level that would have provided a bottle of pills to nearly every adult in the country. That expansion created a vast pool of physiological dependence and addiction, normalised opioid use, and built the demand that heroin and then illicitly manufactured fentanyl would later serve. When medicine reversed course, it often did so abruptly, cutting patients off without offering treatment, and in doing so pushed some of them toward a far more dangerous supply. Both the expansion and the contraction were medical acts. Both carried consequences that the profession did not foresee and, for too long, did not own. To name this history is not to assign the whole epidemic to physicians. Manufacturers, distributors, pharmacies, regulators, and the economics of illicit markets all played their parts, and the chapters that follow give them their due. But clinicians remain the group with the most direct power to change outcomes, because the evidence-based response to opioid use disorder runs through clinical hands: a prescription for buprenorphine, a dose of methadone, a naloxone kit, a referral that actually connects. Understanding how the profession helped make the crisis is the best preparation for understanding how it can help end it. The shape of the book The first two chapters are historical. Chapter 1 traces the origins of the epidemic in the pain-management reforms of the 1980s and 1990s, the marketing of extended-release oxycodone, and the regulatory and commercial failures that allowed prescribing to grow unchecked. Chapter 2 follows the epidemic through its successive waves, from prescription pills to heroin, from heroin to illicit fentanyl, and on to the polysubstance and adulterant-laden supply of the present, ending with the recent decline and the competing explanations for it. Chapter 3 turns to the clinical nature of the condition: the neurobiology of tolerance and dependence, the diagnostic criteria for opioid use disorder, and the epidemiology of mortality that makes it one of the most lethal conditions in medicine. It explains why detoxification without ongoing treatment so often ends in death, a point that underpins every chapter that follows. Chapters 4 and 5 are pharmacological. Buprenorphine receives the closest attention, because its unusual properties as a high-affinity partial agonist explain both its safety and the clinical difficulties of starting it in the fentanyl era. Methadone, the oldest and in some respects the most powerful of the medications, is treated alongside the peculiar regulatory structure in which American law confines it. Chapter 6 covers naloxone and the wider set of harm reduction tools, from syringe services to drug checking, that keep people alive long enough to reach treatment. Chapter 7 examines supervised consumption facilities, now usually called overdose prevention centres in the United States: what the evidence from Sydney, Vancouver, New York, and elsewhere shows about their effects on mortality, public order, and engagement with care, and why they remain legally precarious. Chapter 8 makes the case for the emergency department as the most important single point of entry into treatment, and describes how emergency medicine has moved from treating overdose as an acute poisoning to treating it as a sentinel event in a chronic disease. Chapter 9 turns to policy: prescribing guidelines and their unintended consequences, the regulatory reforms of 2023 and 2024, the settlement money now flowing to states and localities, the treatment of people in jails and prisons, and the contested experiments with decriminalisation. The conclusion asks what the recent decline in deaths does and does not tell us, and what clinicians and health systems should do differently as a result. A note on scope and on doses This is a clinical and policy book written for clinicians, public-health professionals, and serious students. It describes the pharmacology of the medications used to treat opioid use disorder and reverse overdose at the level of mainstream clinical guidance. Where doses and thresholds are given, they reflect widely used guidance at the time of writing and are offered for education, not as orders. Protocols differ between institutions and jurisdictions, product labelling changes, and the drug supply itself shifts faster than guidelines can. Readers who treat patients should defer to their local protocols, current product information, and specialist colleagues. The book describes the illicit drug supply only insofar as clinicians need to understand what their patients are exposed to. It contains nothing about how drugs are obtained, prepared, or used, and it would not be improved by such material. The subject here is how health systems respond. Most of the data and examples are drawn from the United States, which has experienced the most severe opioid epidemic of any high-income country and generated most of the relevant research. Canada, whose experience runs closely parallel and whose policy experiments have been bolder, features throughout, as do Australia and parts of Europe where their experience illuminates a point. The underlying pharmacology and clinical logic travel well; the legal and regulatory details do not, and readers elsewhere should translate accordingly. Finally, a word on language. The book uses person-first terms such as "person with opioid use disorder" and "person who uses drugs" rather than labels such as "addict" or "abuser." This is not etiquette. There is experimental evidence that clinicians who read the word "abuser" are more inclined to favour punitive rather than therapeutic responses, and stigma is one of the barriers this book argues must come down. The same reasoning lies behind the shift from "medication-assisted treatment" toward "medications for opioid use disorder": the medications are not an adjunct to treatment. For most patients, they are the treatment. Chapter 1: How Medicine Made an Epidemic The story of the American opioid epidemic usually begins in 1996, with the launch of OxyContin. That is a reasonable place to put a marker, but it obscures the more uncomfortable fact that the ground had been prepared for more than a decade, largely by clinicians acting in good faith. The expansion of opioid prescribing was not imposed on an unwilling profession by a single company. It was the convergence of a sincere humanitarian reform, a thin and misread evidence base, a regulatory system designed for a different problem, and a commercial strategy that exploited all three. Understanding how each piece fitted together is the precondition for understanding why the crisis proved so hard to reverse. The undertreatment of pain and the reform that followed By the early 1980s, there was solid evidence that pain was badly managed in American hospitals. Studies of postoperative patients and of people dying of cancer repeatedly found that large proportions experienced severe, unrelieved pain, and that physicians and nurses routinely gave less analgesia than was prescribed or required, often because of an exaggerated fear of addiction. The hospice movement, the World Health Organization's analgesic ladder for cancer pain published in 1986, and a generation of pain specialists made the case that opioids were being withheld from people who needed them, out of prejudice rather than pharmacology. They were largely right about cancer pain and acute pain in hospital. The difficulty arose when the argument was extended to chronic non-cancer pain: back pain, osteoarthritis, fibromyalgia, headache, and the many other conditions that bring patients to primary care month after month. Here the evidence was thin. An influential 1986 paper by Russell Portenoy and Kathleen Foley described 38 patients with chronic non-malignant pain treated with opioids, most of whom were said to have done well without developing problematic use. It was a small case series, candidly presented as such, but it became a foundation stone for a broader movement arguing that long-term opioid therapy was safe and effective for chronic pain in carefully selected patients. Even more consequential was a five-sentence letter to the editor published in the New England Journal of Medicine in January 1980. Jane Porter and Hershel Jick, working with a hospital drug surveillance database, reported that among nearly 12,000 hospitalised patients who had received at least one opioid, they found only four documented cases of addiction in patients without a prior history. The letter said nothing about outpatients, nothing about long-term use, and nothing about chronic pain. It was a note on inpatient monitoring. Yet over the following decades it was cited hundreds of times, often as evidence that addiction was "rare" among patients treated with opioids. When a team of Canadian researchers traced its citation history in 2017, they found more than 600 citing articles, the large majority of which described it as evidence that addiction was rare, and few of which noted that it concerned only hospitalised patients. Jick himself later said he was mortified by how the letter had been used. The lesson is not that pain specialists were reckless. Most were motivated by real suffering they saw every day. The lesson is about the epistemics of clinical reform. A movement that is morally right in its core claim (pain is undertreated) is vulnerable to accepting weak evidence for its extensions (therefore opioids are safe for long-term use in chronic pain), because the evidence feels confirmatory and the critics seem to be on the wrong side of a moral question. The same dynamic recurs in later chapters, in the opposite direction, when the reaction against opioids produced its own overreach. Pain as the fifth vital sign By the mid-1990s the reform had become institutional. The American Pain Society promoted the idea of pain as "the fifth vital sign," to be assessed and recorded alongside temperature, pulse, respiration, and blood pressure. The Veterans Health Administration adopted the principle in 1999. In 2001 the Joint Commission, which accredits American hospitals, introduced standards requiring organisations to assess and manage pain. The standards did not instruct anyone to prescribe opioids, and the Commission has since emphasised that point, but in practice they coincided with a period when pain scores became metrics to be improved and opioids were the most reliable way to improve them quickly. Patient satisfaction surveys added a further pressure. When hospital reimbursement and physician evaluations began to incorporate questions about whether pain had been well controlled, the incentives tilted further toward liberal prescribing. Few individual clinicians made a conscious decision to prescribe more opioids for commercial or reputational reasons, but the system around them rewarded it, and the default shifted. The numeric pain scale itself contributed to the drift. A patient's report of pain as seven out of ten became a target to drive down, and the measure carried no information about function, about the natural course of the underlying condition, or about the risks of treatment. A generation of clinicians was trained to treat the number. Opioids reduce the number in the short term more reliably than almost anything else, which made them the obvious choice in a system that counted only the number. OxyContin and the marketing of a molecule Into this environment Purdue Pharma launched OxyContin, an extended-release formulation of oxycodone approved by the Food and Drug Administration in December 1995 and marketed from 1996. Oxycodone itself was not new; it had been used for decades, usually combined with paracetamol in short-acting tablets. What was new was the formulation and the campaign. The extended-release mechanism was designed to deliver oxycodone steadily over twelve hours. The FDA-approved label at the time included a statement that the delayed absorption was believed to reduce the abuse liability of the drug. That sentence, which rested on theory rather than evidence, became the core of Purdue's message to prescribers: that OxyContin was less likely to be misused than short-acting opioids, and that the risk of addiction in pain patients was very low. Sales representatives, incentivised with substantial bonuses, carried that message to tens of thousands of physicians, with particular attention to the highest-volume prescribers, many of them in primary care. Purdue funded pain education programmes, speakers' bureaus, and patient advocacy groups that promoted opioid therapy for chronic pain. It introduced the concept of "pseudoaddiction," which held that drug-seeking behaviour in pain patients was often a sign of undertreated pain that called for higher doses. The pharmacology did not support the marketing. Twelve-hour dosing frequently wore off early, leading patients to experience withdrawal symptoms and craving between doses, and the company encouraged dose increases rather than more frequent dosing to address this. The tablets contained far more oxycodone than the short-acting combination products, and crushing them defeated the release mechanism, delivering the full dose at once. Early reports of misuse came from rural areas in Appalachia, Maine, and elsewhere within a few years of launch. OxyContin sales rose from roughly 48 million dollars in 1996 to more than a billion dollars by 2000. In 2007 the Purdue Frederick Company, an affiliate, pleaded guilty to a federal felony charge of misbranding OxyContin with intent to defraud or mislead, and three executives pleaded guilty to misdemeanour misbranding; the company and the executives paid around 630 million dollars. The penalty was modest relative to revenue, and prescribing continued to rise. In 2020 Purdue itself pleaded guilty to three federal felonies, including conspiracy to defraud the United States and violations of anti-kickback law, as part of a resolution valued at more than eight billion dollars, most of it a claim in bankruptcy that would never be paid in full. It would be a mistake to make OxyContin the whole story. Other manufacturers marketed their own long-acting opioids aggressively. Hydrocodone combination products, which were not branded blockbusters, were consistently the most commonly prescribed opioids in the United States for most of this period. What Purdue did, more than any other actor, was to change the professional consensus about the risk of addiction. Once that consensus had shifted, prescribing expanded across products and specialties. Distribution, dispensing, and the failure of controls Opioids are controlled substances under federal law, and every pill passes through a supply chain that is in principle tightly monitored: manufacturers, wholesale distributors, pharmacies, and prescribers all register with the Drug Enforcement Administration and are required to report and guard against diversion. The epidemic exposed how poorly that system functioned when the commercial incentives of every participant pointed in the same direction. Distributors were obliged to identify and report suspicious orders. Investigative reporting and litigation later documented instances in which small pharmacies in towns of a few thousand people received millions of hydrocodone and oxycodone pills over a few years, without effective intervention. In Florida in the late 2000s, so-called pill mills, cash-only clinics that dispensed opioids with minimal examination, drew customers from across the eastern United States. Florida's crackdown in 2010 and 2011, which included prescription drug monitoring and new rules on pain clinics, was followed by measurable reductions in prescribing and in prescription-opioid deaths in the state, an early demonstration that supply-side regulation could change prescribing quickly. The three largest drug distributors, McKesson, Cardinal Health, and AmerisourceBergen (now Cencora), together with Johnson & Johnson, agreed in 2021 and 2022 to national settlements totalling about 26 billion dollars over eighteen years, resolving thousands of claims brought by states and local governments. Pharmacy chains and other manufacturers followed with their own settlements. The consulting firm McKinsey & Company, which had advised Purdue on how to "turbocharge" OxyContin sales, settled with states for more than 570 million dollars in 2021 and later reached a separate federal criminal resolution. The total value of opioid settlements now exceeds 50 billion dollars, a sum whose uses are discussed in Chapter 9. Purdue's own resolution took years longer. A bankruptcy plan that would have shielded members of the Sackler family from civil liability was struck down by the United States Supreme Court in June 2024, in Harrington v. Purdue Pharma, on the ground that bankruptcy law did not authorise such non-consensual releases of third parties. A revised plan, in which the Sacklers' contribution rose to about 6.5 billion dollars within a total of at least 7.4 billion, was confirmed by the bankruptcy court in November 2025. Purdue is to be succeeded by a new company owned by a charitable foundation. The scale of exposure The result of all these forces was an enormous increase in the quantity of opioids prescribed. The rate of opioid prescriptions dispensed in the United States peaked in 2012 at more than 81 prescriptions for every 100 people, according to the CDC. Morphine milligram equivalents per capita rose roughly fourfold from the late 1990s to the peak. Prescriptions for longer durations and higher doses grew particularly fast, and those are precisely the prescriptions most strongly associated with dependence, opioid use disorder, and overdose. The epidemiology of risk from prescribed opioids has since been worked out in some detail. The likelihood of continued use one year after an initial prescription rises with the number of days supplied in the first prescription, with a noticeable increase after only a few days. Risk of overdose rises steeply with daily dose, and concurrent prescription of benzodiazepines multiplies it. Many people who received opioids after surgery never used most of their pills, and those leftover supplies in medicine cabinets became a principal source for adolescents and young adults beginning non-medical use. National surveys repeatedly found that most people who misused prescription pain relievers obtained them from friends or relatives, not from a physician directly. Not every person exposed developed a problem; the great majority did not. But when the denominator is tens of millions of people, a small percentage is a very large number. By the end of the 2000s, several million Americans met criteria for a prescription opioid use disorder, and the death rate from prescription opioid overdose had more than tripled since 1999. The parallel in Canada and the limits of the American exception It is tempting to treat the epidemic as a peculiarly American pathology, explained by direct-to-physician marketing, fragmented insurance, and weak regulation. The Canadian experience complicates that account in a useful way. Canada has a single-payer system for physician and hospital services, tighter controls on pharmaceutical advertising, and a smaller and more concentrated pharmaceutical market. It nonetheless became, by the late 2000s, one of the highest per-capita consumers of prescription opioids in the world, second only to the United States on several measures. OxyContin was marketed in Canada with many of the same claims, and Canadian provinces saw the same pattern of rising prescription opioid deaths in the 2000s, followed by the arrival of illicit fentanyl in the mid-2010s, which hit British Columbia and Alberta first and hardest. The comparison suggests that the professional consensus about opioid safety, more than any single feature of American health financing, drove the expansion. A belief that spread through medical education, specialty societies, and continuing professional development crossed national borders easily. By contrast, countries in western Europe and East Asia, where cultural and regulatory attitudes to opioids remained more conservative and where pain reform focused more narrowly on cancer and palliative care, did not experience comparable prescription-driven epidemics, although several now face their own problems with illicit synthetic opioids. The comparison also illuminates the later chapters of the story. Canada's response diverged from the American one in important ways: it moved earlier and further toward supervised consumption, safer supply, and decriminalisation, while also permitting methadone to be prescribed by trained physicians in primary care and dispensed in community pharmacies. Its overdose death rates in the worst-affected provinces have nonetheless remained high. The divergence offers something close to a natural experiment, imperfect as all such comparisons are, and later chapters draw on it. Prescribing after the reckoning The period of excess left behind practices that clinicians are still working to correct, and some of the reforms are worth naming because they show what responsible prescribing looks like now. Surgical teams in Michigan, through a statewide quality collaborative, developed procedure-specific recommendations for the number of opioid tablets to prescribe after common operations, based on how many patients actually used; prescribing fell substantially without any increase in reported pain or requests for refills. Similar evidence-based prescribing guides have spread to other states and specialties. Dentistry, once a major source of first opioid exposures for adolescents having wisdom teeth removed, has shifted toward combinations of ibuprofen and paracetamol, which perform as well or better than opioids for most dental pain. Prescription drug monitoring programmes, once optional and cumbersome, are now integrated into electronic health records in many states, and checking them before prescribing a controlled substance is mandatory in most. Safe disposal programmes, including pharmacy take-back kiosks, aim to reduce the reservoir of leftover tablets in household cabinets. Each of these measures is modest. Together they have helped bring per-capita opioid prescribing in the United States back to levels last seen in the early 2000s, though still above those of most comparable countries. What "iatrogenic" does and does not mean There is a temptation, having told this story, to frame the whole epidemic as a morality play in which greedy companies corrupted gullible doctors. Parts of that frame are justified, and the courts have confirmed them. But it is incomplete in ways that matter for clinical practice today. First, many clinicians who prescribed generously were responding to real patients in real pain, often with few alternatives. Physiotherapy, psychological treatment for chronic pain, and multidisciplinary pain programmes were frequently unavailable or not covered by insurance, while a prescription took a minute and was reimbursed. The structure of American health care made opioids the path of least resistance. That structure has not entirely changed. Second, the iatrogenic phase established not just addiction but also physiological dependence in large numbers of people who did not have opioid use disorder at all: patients on stable long-term therapy who functioned well and whose principal risk would later come from the system's attempt to undo what it had done. Chapter 9 describes how abrupt tapering and discontinuation, encouraged by a misreading of the 2016 prescribing guideline, caused documented harm to this group. Third, and most importantly for the rest of this book, the profession that created the demand had in its hands, throughout, the remedies that could have limited the damage. Methadone had been in use as a maintenance treatment since the 1960s. Buprenorphine was approved for office-based treatment of opioid dependence in 2002. Naloxone had been used to reverse overdose since the 1970s. Each was underused, for reasons that had little to do with evidence. The first iatrogenic failure was overprescribing. The second, longer-lasting failure was the reluctance to treat the disorder that overprescribing had helped produce. That second failure is the more instructive, because it is still with us. The next chapter follows the epidemic as it moved beyond the pharmacy into the illicit market, where the cost of that reluctance was counted in hundreds of thousands of lives. Chapter 2: From Pills to Fentanyl Epidemiologists describe the American overdose epidemic in waves, and the description has stuck because it captures something real. Each wave was driven by a different class of drug, killed a somewhat different population, and demanded a different response. But the metaphor can mislead. Waves suggest that one recedes as the next arrives. In practice each wave was layered on top of the one before, the population at risk carried forward from phase to phase, and the drug supply grew steadily more lethal. The clinical consequence is that a person who began with a prescription in 2005 might, by 2020, be using a supply that bore no resemblance to anything a physician had ever prescribed. The first wave: prescription opioids, 1999 to 2010 The first wave, conventionally dated from 1999, when consistent national mortality coding begins, was dominated by deaths involving prescription opioids: oxycodone, hydrocodone, methadone prescribed for pain, and others. Deaths rose steadily through the 2000s. The people dying were older on average than in later waves, more often white, and disproportionately from rural areas and from states such as West Virginia, Kentucky, and Utah, where prescribing was heaviest. Many deaths involved combinations of opioids with benzodiazepines or alcohol. Methadone deserves a specific note here because it is otherwise discussed in this book as a treatment. In the 2000s, methadone was increasingly prescribed for pain, partly because it was cheap and on many insurers' preferred lists. Its long and variable half-life, which makes it effective as a once-daily maintenance medication in a supervised setting, made it dangerous when prescribed for pain by clinicians unfamiliar with its accumulation. Methadone prescribed for pain accounted for a disproportionate share of prescription opioid deaths relative to its share of prescriptions, and later policy restricted its use for that purpose. Deaths among patients in supervised opioid treatment programmes, by contrast, were and remain relatively uncommon. The first wave prompted the first serious regulatory response. States built prescription drug monitoring programmes, databases that allow prescribers and pharmacists to see a patient's history of controlled-substance prescriptions. Some states regulated pain clinics. In 2010 Purdue introduced an abuse-deterrent reformulation of OxyContin that was harder to crush or dissolve. These measures, combined with growing clinical awareness, slowed prescribing growth, and prescription opioid deaths began to plateau around 2011. The second wave: heroin, 2010 to 2013 From 2010, deaths involving heroin rose sharply, more than tripling in the next five years. The timing was not coincidental. Several careful studies examined what happened after the OxyContin reformulation and found that, in states where non-medical use of OxyContin had been most common before the change, heroin overdose deaths subsequently rose more steeply. The reformulation reduced misuse of OxyContin itself, but a substantial proportion of the people who had been misusing it moved to heroin, which was cheaper and widely available. This is the first clear instance of a pattern that recurs throughout the epidemic: interventions that restrict supply without providing treatment displace demand rather than eliminating it. A person with an established opioid use disorder does not stop needing opioids because a particular formulation has become harder to misuse. Unless treatment is offered at the moment access is restricted, the predictable result is migration to a riskier source. Surveys conducted at the time supported the displacement account from the other side. Among people entering treatment for heroin use disorder in the 2000s, a large majority reported that their first regular opioid had been a prescription drug, a reversal of the pattern in the 1960s, when most began with heroin. The heroin users of the second wave were younger than those of the first, and increasingly from suburban and rural areas that had little previous experience of heroin. The third wave: illicitly manufactured fentanyl, 2013 onward Fentanyl is a synthetic opioid, first synthesised in 1960 and used in anaesthesia and severe pain for decades. It is roughly fifty to one hundred times more potent than morphine by weight. Pharmaceutical fentanyl, used in hospitals and as transdermal patches, played only a minor part in the epidemic. What arrived from around 2013 was illicitly manufactured fentanyl, produced outside any regulated system and introduced into the heroin supply, first in parts of New England and the mid-Atlantic and later across the country. The change was transformative. Within a few years, in many eastern markets, what was sold as heroin was predominantly or entirely fentanyl. By the late 2010s, synthetic opioids other than methadone, a category dominated by illicit fentanyl and its analogues, were involved in the majority of overdose deaths in the United States. Fentanyl reached western states later, but by the early 2020s it was nationwide. Three properties of illicit fentanyl explain its lethality. The first is potency: because a lethal quantity is so small, minor variation in how it is distributed through a batch can produce enormous differences in the dose a person receives. The second is rapid onset: fentanyl is highly lipophilic and reaches the brain within minutes, so that respiratory depression can develop before bystanders recognise what is happening. The third is unpredictability: people buying drugs on the illicit market usually do not know what they are getting, and an increasing share of deaths involved people who believed they were taking something else, including counterfeit pills made to resemble oxycodone or benzodiazepines. Fentanyl also changed the clinical picture of opioid use disorder. People with regular exposure to fentanyl develop a high degree of tolerance, and because the drug is so lipophilic it is stored in fatty tissues and released slowly. This has practical consequences for starting buprenorphine, which Chapter 4 describes in detail: the old rules of thumb about how long to wait after the last use before the first dose became unreliable. It also affected naloxone, since more potent opioids may require more than one dose to reverse, a point taken up in Chapter 6. The fourth wave: stimulants, sedatives, and a mixed supply By the late 2010s a further shift was visible. A growing proportion of overdose deaths involved stimulants, methamphetamine and cocaine, usually in combination with fentanyl. Some of this reflected intentional use of both; some reflected contamination of the stimulant supply with fentanyl. Researchers began to describe a fourth wave characterised by polysubstance use, in which the boundaries between opioid and stimulant epidemics broke down. The supply has also acquired non-opioid adulterants. Xylazine, a veterinary sedative and alpha-2 adrenergic agonist, spread through the fentanyl supply in Philadelphia and then across the eastern United States from around 2019. Its sedative effects are not reversed by naloxone, it is associated with severe, slow-healing skin wounds, and people with regular exposure can experience a distinct withdrawal syndrome. In April 2023 the White House designated fentanyl combined with xylazine an emerging threat to the nation, the first time that designation had been used. More recently, medetomidine, another veterinary alpha-2 agonist considerably more potent than xylazine, has appeared in the supply in Philadelphia, Chicago, and elsewhere. Clinicians in emergency departments and hospitals have reported a severe withdrawal syndrome associated with it, featuring marked agitation, hypertension, and tachycardia, that can be difficult to manage and sometimes requires intensive care. Novel synthetic opioids, including the nitazenes, a class of potent benzimidazole opioids, have also been detected intermittently in North America and Europe. For clinicians the lesson of the fourth wave is humility about what a patient has actually been exposed to. A person presenting after an overdose may have taken fentanyl, a nitazene, a sedative adulterant, a stimulant, or several at once. Urine drug screens in common use often detect none of the adulterants. Local drug-checking data and public-health alerts have become as relevant to clinical practice as published guidelines. Table 1 summarises the four waves on the dimensions most relevant to clinical and policy response. Table 1. The four phases of the US overdose epidemic. Phase Onset Main drugs Most affected Typical response First 1999 Prescribed opioids, benzodiazepines Older, rural adults Monitoring, clinic rules Second 2010 Heroin Younger adults Naloxone, early treatment Third 2013 Illicit fentanyl Broad; rising in Black and Indigenous people Buprenorphine, harm reduction Fourth 2016 on Fentanyl, stimulants, sedatives Older Black men, Indigenous people Drug checking, ED and jail treatment The shifting demography of death The early epidemic was often described as a crisis of white, rural, and suburban America, and media coverage reinforced that framing. It was always an oversimplification, and by the fentanyl era it had become plainly wrong. From around 2015, overdose death rates rose fastest among Black Americans, and by 2020 the rate among Black Americans exceeded that among white Americans. American Indian and Alaska Native people have experienced the highest death rates of all in recent years. Overdose mortality among older Black men in particular, many of them people who had used heroin for decades and survived earlier eras, rose steeply as fentanyl replaced heroin in urban markets. These shifts matter because the treatment system expanded unequally. Studies through the late 2010s found that buprenorphine prescriptions went disproportionately to white patients with private insurance or the ability to pay cash, while Black patients with opioid use disorder were more likely to receive methadone, with its daily clinic attendance, or no medication at all. A policy response built around the early, whiter face of the epidemic was poorly matched to the people dying in the third and fourth waves. Reading the numbers Because so much of the policy debate turns on mortality statistics, clinicians and public-health professionals benefit from understanding how those statistics are produced and where they mislead. National overdose counts in the United States come from death certificates, compiled by the CDC's National Center for Health Statistics. Determining that a death was caused by an overdose, and which drugs were involved, requires a medicolegal death investigation and toxicology testing, which can take months. The CDC therefore publishes provisional counts, adjusted for expected delays in reporting, and updates them as records are completed. The figures for any recent period are estimates that typically rise somewhat as late records arrive, and final data lag by about a year. Drug involvement is recorded using codes for every substance mentioned on the death certificate. A single death involving fentanyl, cocaine, and xylazine is counted under each of those categories, so the drug-specific totals sum to more than the total number of deaths. The quality of drug-specific reporting depends on the thoroughness of toxicology and on the investigator's practice of listing specific drugs rather than writing "drug overdose" unspecified, and this varies considerably between jurisdictions. Comparisons between states, especially for less commonly tested substances, should be read with that in mind. Non-fatal overdoses are counted less reliably still. Emergency department syndromic surveillance, emergency medical services records of naloxone administration, and hospital discharge data each capture part of the picture, and many overdoses reversed by bystanders never reach any data system. A fall in emergency department visits for overdose could mean fewer overdoses, or more reversals in the community, or more people declining transport to hospital. Finally, rates matter as much as counts. A state with a small population may have few deaths but a very high rate, and national figures can conceal large differences by age, race, and region. The most useful data for clinical planning are usually local: county or city surveillance, drug-checking results, and the experience of emergency departments and harm reduction services in the area. The peak and the decline Overdose deaths in the United States rose almost every year from 1999 to 2023, with an especially sharp acceleration in 2020 and 2021, during the COVID-19 pandemic, when treatment and harm reduction services were disrupted, social isolation increased the chance of using alone, and the drug supply became more volatile. Provisional CDC data show deaths peaking at roughly 110,000 to 112,000 in the twelve-month periods ending in mid-2023. Then the numbers fell, quickly and broadly. The CDC's provisional estimates put the total at about 81,300 deaths in 2024 and about 70,000 in 2025, a decline of almost 14 percent in the most recent year and the third consecutive annual decrease. Opioid-involved deaths fell from roughly 55,300 to roughly 44,600, close to a fifth. Deaths involving cocaine and methamphetamine also fell. The declines were large in some states, more than 25 percent in a year in Rhode Island, New York, North Carolina, Alabama, and Vermont, while a few western states, including New Mexico, Arizona, and Colorado, saw increases. There is no consensus explanation for the decline, and honest accounts acknowledge several contributors whose relative weight is uncertain. The candidates include the following. Changes in the drug supply. Some analysts point to evidence that illicit fentanyl became less pure in parts of the country from around 2023, possibly reflecting enforcement pressure on precursor chemicals and on trafficking organisations. If the average potency of a dose fell, fewer doses would be fatal. This hypothesis is plausible but difficult to verify with existing surveillance data. Expanded naloxone. Naloxone became available without a prescription in March 2023, and distribution through community programmes, pharmacies, schools, and public venues increased substantially. More reversals mean fewer deaths, although naloxone alone cannot explain why deaths fell faster in some places than others. Expanded treatment access. The removal of the federal waiver requirement for buprenorphine in early 2023 and the methadone regulatory reforms of 2024, described in Chapters 4 and 5, broadened who could prescribe and on what terms. The effects on the number of people treated have been more modest than advocates hoped, but expansion in emergency departments, jails, and telehealth has continued. Depletion of the susceptible population. The grimmest hypothesis is that so many people at high risk have already died that the pool of those most vulnerable has shrunk. Some demographic analyses lend partial support, showing that cohorts entering adulthood more recently have lower rates of opioid initiation than those of the 2000s. Changes in patterns of use. There are reports of shifts in route of administration, from injection toward smoking, which may carry a lower risk of fatal overdose, and of changes in how people who use drugs manage their own risk, including not using alone and using test strips. The most defensible reading is that several of these forces operated together, and that their relative contribution differs by region. That uncertainty is itself a reason for caution. A decline driven largely by supply changes could reverse if the supply shifts again; one driven by expanded treatment and naloxone is more durable. The rise in a few western states and the persistence of very high death rates among Black and Indigenous Americans, whose rates fell less than the national average in some analyses, are reminders that the epidemic is not over. What can be said with confidence is that the treatments and harm reduction tools described in the following chapters have strong individual-level evidence of preventing death, and that the populations still dying are, overwhelmingly, people who were not engaged with those tools at the time. The next chapter explains why opioid use disorder is so lethal and why the interventions that work are the ones that keep people continuously connected to care. Chapter 3: Opioid Use Disorder as a Chronic Lethal Illness Every successful treatment for opioid use disorder rests on a premise that much of the health system has been slow to accept: that the condition is a chronic illness with a high case-fatality rate, not a moral failing, a lifestyle choice, or a problem to be solved by a few weeks of abstinence. The premise is not a rhetorical position. It follows from what is known about the neurobiology of opioids, from the natural history of the disorder, and above all from the mortality data. This chapter sets out that evidence, because it is the foundation for every clinical choice that follows. Tolerance, dependence, and addiction Opioids act principally at the mu-opioid receptor, a G-protein-coupled receptor distributed throughout the central nervous system and elsewhere. Activation of the receptor produces analgesia, euphoria, sedation, constipation, miosis, and, critically, depression of the brainstem respiratory centres, which reduces the ventilatory response to rising carbon dioxide. Death from opioid overdose is almost always death from respiratory depression: breathing slows, then stops, and the brain is injured by hypoxia within minutes. With repeated exposure, three related but distinct phenomena develop. They are often confused, and the confusion has caused real harm. Tolerance is the reduction in effect of a given dose over time, so that higher doses are needed to produce the same response. It develops at different rates for different effects. Tolerance to euphoria and to respiratory depression develops relatively quickly; tolerance to constipation and miosis develops slowly or hardly at all. Importantly, tolerance is lost within days to weeks of stopping, a fact with lethal implications discussed below. Physiological dependence is the state in which abrupt cessation or administration of an antagonist produces a withdrawal syndrome. Opioid withdrawal is characterised by anxiety, restlessness, yawning, lacrimation, rhinorrhoea, piloerection, mydriasis, muscle and bone aches, abdominal cramps, vomiting, and diarrhoea. It is profoundly unpleasant but, in otherwise healthy adults, rarely life-threatening in itself. Dependence is an expected physiological adaptation to sustained opioid exposure. A patient on stable long-term opioid therapy for pain is dependent; so is a patient maintained on methadone or buprenorphine. Dependence alone is not a disorder. Addiction, clinically termed opioid use disorder, is a behavioural syndrome: compulsive use despite harm, loss of control, craving, and the progressive narrowing of life around obtaining and using the drug. It reflects durable changes in brain circuits involved in reward, motivation, stress, and executive control. Repeated exposure to opioids, in people with susceptibility shaped by genetics, early adversity, mental illness, and social circumstance, alters dopaminergic signalling in the mesolimbic pathway, strengthens learned associations between drug cues and reward, and recruits stress systems so that the absence of the drug becomes aversive. Over time, use shifts from being driven by the pursuit of pleasure toward being driven by the avoidance of withdrawal and dysphoria. The practical point is that dependence and addiction can occur independently. Confusing them led, in the 1990s, to the pseudoaddiction concept, which treated signs of addiction as evidence of undertreated pain. Confusing them in the opposite direction, in the late 2010s, led some clinicians to treat every dependent pain patient as if they had a use disorder, and to taper them abruptly. Neither error is benign. Diagnosis The fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5, and its 2022 text revision) defines opioid use disorder by eleven criteria occurring within a twelve-month period. They fall into four groups: impaired control (taking more or for longer than intended, unsuccessful efforts to cut down, much time spent obtaining or recovering, craving); social impairment (failure to fulfil obligations, continued use despite social problems, giving up important activities); risky use (use in hazardous situations, continued use despite physical or psychological harm); and pharmacological criteria (tolerance and withdrawal). Two or three criteria indicate mild disorder, four or five moderate, and six or more severe. A crucial qualification is built into the criteria. Tolerance and withdrawal are not counted toward the diagnosis when a person is taking opioids solely under appropriate medical supervision. This reflects the distinction between dependence and addiction. A patient on prescribed opioids who is physiologically dependent but shows none of the behavioural criteria does not have opioid use disorder. In clinical practice the diagnosis is usually not difficult once someone thinks to ask. The difficulty is that many clinicians do not ask, or ask in ways that invite denial. Brief validated screening questions, such as the single question about how many times in the past year a person has used an illegal drug or a prescription medication for non-medical reasons, are effective as a first step. A non-judgemental conversation that begins from the patient's goals generally yields more accurate information than a checklist delivered with visible disapproval. The natural history of the disorder Opioid use disorder typically follows a relapsing and remitting course over many years. Long-term cohort studies of people treated for heroin dependence, some following participants for more than three decades, show that a proportion achieve sustained remission, often after multiple episodes of treatment, while others cycle in and out of use and a substantial fraction die prematurely. Remission becomes more stable the longer it lasts, but relapse after years of abstinence is not rare. This pattern resembles other chronic illnesses. Comparisons drawn by addiction researchers in the late 1990s and since show that rates of non-adherence and relapse in substance use disorders are similar to those for hypertension, asthma, and type 2 diabetes. Nobody concludes from a diabetic patient's raised HbA1c that insulin has failed or that the patient should be discharged from care. The same reasoning applies to a person with opioid use disorder who uses again while on treatment: it is a signal to adjust the treatment, not a justification for ending it. Mortality: the central fact The single most important fact about untreated opioid use disorder is how often it kills. Cohort studies of people who inject drugs or who are dependent on opioids have consistently found all-cause mortality ten or more times higher than in the general population of the same age. The principal causes are overdose, infections (HIV, hepatitis C, endocarditis, and other injection-related infections), trauma, and suicide. Two large bodies of evidence show what treatment does to that risk. The first is a systematic review and meta-analysis of cohort studies published in the BMJ in 2017 by Luis Sordo and colleagues, which pooled data on more than 122,000 people treated with methadone and nearly 16,000 treated with buprenorphine. All-cause mortality among people receiving methadone was 11.3 per 1,000 person-years in treatment and 36.1 per 1,000 person-years out of treatment, a ratio of about three. For buprenorphine the corresponding rates were 4.3 and 9.5. Overdose mortality showed similar or larger differences. The analysis also identified two periods of particular danger: the first four weeks of methadone treatment, when doses are being increased and tolerance is uncertain, and the first four weeks after leaving treatment of either kind. The second comes from a linked-data cohort study in Massachusetts, published in the Annals of Internal Medicine in 2018 by Marc Larochelle and colleagues. They followed 17,568 adults who had survived an opioid overdose. In the following year, only a minority received any medication for opioid use disorder. Compared with no medication, methadone was associated with an approximately 59 percent reduction in opioid-related mortality and buprenorphine with an approximately 38 percent reduction. Naltrexone, an antagonist, showed no association with reduced mortality, though the number of patients receiving it was small and treatment durations were short. Table 2 sets these figures alongside one another. Table 2. Mortality in and out of medication treatment for opioid use disorder. Source Population Comparison Finding Sordo et al., BMJ 2017 122,885 on methadone (cohort meta-analysis) All-cause mortality in vs out of treatment 11.3 vs 36.1 per 1,000 person-years Sordo et al., BMJ 2017 15,831 on buprenorphine All-cause mortality in vs out of treatment 4.3 vs 9.5 per 1,000 person-years Larochelle et al., Ann Intern Med 2018 17,568 overdose survivors, Massachusetts Opioid-related mortality, methadone vs none Adjusted HR 0.41 Larochelle et al., Ann Intern Med 2018 Same cohort Opioid-related mortality, buprenorphine vs none Adjusted HR 0.62 Larochelle et al., Ann Intern Med 2018 Same cohort Opioid-related mortality, naltrexone vs none No significant association These are observational studies, and observational studies can be confounded: people who stay in treatment may differ in many ways from people who leave. But the effect is large, consistent across countries and decades, biologically plausible, and supported by trials showing that agonist treatment reduces illicit opioid use and retains people in care. Few interventions in medicine reduce all-cause mortality in a high-risk population by half. Agonist treatment for opioid use disorder is one of them. Why detoxification alone is dangerous If the mortality data establish one clinical principle, it is that withdrawal management on its own, often called detoxification, is not a treatment for opioid use disorder, and can be more dangerous than no intervention at all. The reason is tolerance. A person who has been using opioids regularly has a high tolerance to respiratory depression. After a week or two of abstinence, whether in a detoxification unit, a hospital, a jail, or a residential programme without medication, that tolerance falls substantially. Craving and the conditioned associations that drive use, however, persist for months or years. If the person returns to use at a dose they previously tolerated, the result may be fatal. The risk is particularly acute in the fentanyl era, when the dose in any given purchase is unpredictable. The data on release from prison illustrate the danger starkly. A widely cited study of people released from Washington State prisons, published in the New England Journal of Medicine in 2007, found that in the first two weeks after release the risk of death from overdose was more than one hundred times that of the general population. Similar findings have been reported in England, Australia, and elsewhere. The same logic applies to discharge from abstinence-based residential treatment and from hospital. The clinical implication is that when a person with opioid use disorder is offered withdrawal management, it should be offered as a bridge to ongoing medication, not as an end in itself, and that anyone leaving a period of enforced abstinence should receive naloxone and an offer of treatment. A programme that detoxifies a patient and discharges them without medication or naloxone has, in effect, made them more vulnerable to overdose than when they arrived. Co-occurring conditions Opioid use disorder rarely travels alone. A large proportion of patients have co-occurring mental illness, most commonly depression, anxiety disorders, post-traumatic stress disorder, and other substance use disorders, especially involving stimulants, alcohol, and benzodiazepines. Chronic pain is common, and pain and opioid use disorder interact in complex ways: pain can drive use, opioid withdrawal increases pain sensitivity, and long-term opioid exposure can itself produce hyperalgesia. Infectious complications are a major source of morbidity. Hepatitis C infection is endemic among people who inject drugs, and the availability since 2014 of direct-acting antiviral regimens with cure rates above 95 percent has made treatment straightforward; there is no longer any justification for withholding it from people who are actively using. Injection-related bacterial infections, including endocarditis, osteomyelitis, and skin and soft tissue infections, have risen sharply in the fentanyl era. Patients admitted to hospital for these infections who are not started on medication for opioid use disorder frequently leave before completing treatment, often because untreated withdrawal and hospital policies make staying intolerable, and their readmission and mortality rates are high. Addiction consultation services in hospitals, discussed in Chapter 8, have been shown to increase the start of medication and to reduce such departures. Housing instability, incarceration, poverty, and social isolation shape both the risk of the disorder and the chance of recovery. Treatment that ignores them is less effective. But they are not prerequisites for medication. One of the more damaging habits of the treatment system has been to require patients to achieve stability in their lives before receiving the medications that would make stability possible. Stigma as a clinical variable It is tempting to treat stigma as a matter of manners, separate from the hard science of pharmacology and mortality. The evidence suggests otherwise. Stigma operates as a clinical variable, shaping who is diagnosed, who is offered treatment, and who stays. Patients with opioid use disorder report, with striking consistency, experiences of being treated with suspicion or contempt in health settings: pain dismissed as drug-seeking, delays in care, comments from staff, and discharge practices that seem punitive. Many delay seeking care as a result, presenting later with more advanced infections or not at all. Experimental work by John Kelly and colleagues at Harvard found that clinicians who read a vignette describing a patient as a "substance abuser" were more likely to judge the patient personally culpable and to favour punitive measures than those who read an identical vignette describing a person "with a substance use disorder." Language shapes judgement even among trained professionals. Stigma also attaches to the treatments themselves. The belief that methadone and buprenorphine merely "substitute one addiction for another" remains widespread among the public, among some professionals in the addiction field, and among patients' families. It reflects the confusion between dependence and addiction described at the start of this chapter. A patient taking a stable daily dose of buprenorphine is physiologically dependent on it, as a patient with hypothyroidism depends on levothyroxine, but shows none of the compulsive, harmful use that defines the disorder. Abstinence-oriented recovery communities, peer-support groups, residential programmes, and drug courts have historically discouraged or prohibited agonist medications, and patients in such settings have sometimes faced pressure to stop. That pressure, given the mortality data, is dangerous. For clinicians, the practical consequences are concrete. Use clinical language. Treat a patient's report of pain as credible until there is reason to think otherwise. Explain clearly that agonist treatment is the standard of care, and support patients who encounter pressure to stop it. And recognise that the trust built in a single respectful encounter may determine whether the patient returns at all. Recovery without a fixed endpoint The concept of recovery has itself been contested. In older usage, recovery meant complete abstinence from all psychoactive substances, sustained over time, often within a twelve-step framework. Contemporary definitions, including the working definition adopted by SAMHSA, describe recovery as a process of change through which people improve their health and wellness, live self-directed lives, and strive to reach their full potential. Under this definition, a person maintained on methadone or buprenorphine can be in recovery, and many are. The distinction matters because it changes what counts as success. Treatment programmes judged by the proportion of patients who become abstinent from all opioids will appear to fail most of the time, and will be tempted to discharge patients who do not progress toward that goal. Programmes judged by survival, retention, reduced illicit use, and improved function will appear, accurately, to succeed for many patients. Most of the barriers to medication described in later chapters have their roots in the older definition. The goals of treatment Taken together, this evidence suggests a clear hierarchy of treatment goals. The first is survival: keeping the person alive through the period of greatest risk. The second is reducing harm, including infection, injury, and the social damage of uncontrolled use. The third is improving function and quality of life, as the person defines them. Abstinence from all opioids, including treatment medications, may be a goal for some patients and is entirely legitimate when they choose it, but it is not the measure of success. A patient who has stopped using illicit fentanyl, is taking buprenorphine daily, and is working and caring for a family is a treatment success, not someone still waiting to recover. The medications that best achieve these goals act at the same receptor as the drugs that cause the disorder. That apparent paradox has been the source of much of the resistance to them. The next two chapters explain why it is not a paradox at all. Hashtags: #TheOpioidCrisis #IatrogenicAddiction #OpioidUseDisorder #HarmReduction #PolicyReform #PrescriptionOpioids #OxyContin #OpioidPrescribing #IllicitFentanyl #PolysubstanceUse #Buprenorphine #Methadone #Naloxone #MedicationsForOpioidUseDisorder #OverdosePrevention #SupervisedConsumption #OverdosePreventionCentres #EmergencyDepartmentInitiation #SyringeServices #DrugChecking #PrescriptionDrugMonitoring #PhysiologicalDependence #OpioidTolerance #StigmaReduction #FutureOfOpioidPolicy

Download the Book (PDF): Introduction A kidney is about the size of a fist and weighs roughly 150 grams. Inside it, around a million filtering units sit in a precise relationship to a blood supply that receives something like a fifth of everything the heart pumps. A liver holds on the order of a hundred billion hepatocytes, each of them within a few cell widths of a capillary. A femur carries several times body weight with every stride and quietly tears down and rebuilds its own substance, a little at a time, for as long as its owner lives. None of these organs is just a collection of the right cells. Each is an arrangement: of cells, of the extracellular matrix they sit in, of vessels that feed them, and of the forces that pass through all three. Tissue engineering is the attempt to build such arrangements outside the body, or to coax the body into building them where it otherwise would not. The ambition is old in spirit and young in practice. The field acquired its name in the late 1980s and its manifesto in 1993, when Robert Langer and Joseph Vacanti published a short review in Science titled simply "Tissue Engineering." Their proposal was that living tissue could be made by combining three ingredients: cells, a scaffold for the cells to grow on, and signals that tell the cells what to become. That triad still organizes the field, and it still appears on the first slide of a great many lectures. The triad is a good place to start and a misleading place to stop. Its weakness is that it treats the scaffold as a shape, the cells as a population, and the signals as chemistry. Three decades of work have shown that this misses most of what decides whether an engineered tissue lives or dies. The scaffold is not mainly a shape; it is a transport medium and a mechanical environment. The cells are not mainly a population; they are a demand for oxygen that grows faster than any passive supply can meet. And the most important signals are often not molecules at all. They are forces: the stiffness of the material beneath a cell, the drag of fluid moving past it, the stress it experiences as it is squeezed through a printing nozzle. The argument of this book This booklet argues one thing: that building replacement tissue is, at bottom, a problem of designing transport and mechanics, and that the three technologies most often discussed in the field — porous ceramic scaffolds, perfusion bioreactors and bio-ink printing — succeed or fail according to how well they manage the physics the cells require. Mass transport, above all the delivery of oxygen, sets the hardest limit. Fluid shear stress is both a hazard and an instruction. Rheology decides whether a bio-ink can be printed at all and whether the cells inside it survive the trip. Stem cells and decellularized matrices bring biology that no engineer could design from scratch, but they do not escape these constraints; they only change the terms on which the constraints must be met. Put more bluntly: most engineered tissues that fail do not fail because their designers chose the wrong growth factor. They fail because their centres starve, because their vessels do not connect, because their matrix is too weak or too stiff, or because their cells were damaged before culture began. The successes of the field, real but narrower than its early promises, are concentrated in tissues that are thin, avascular, or mechanically forgiving: skin substitutes, cartilage, the bladder wall, an acellular blood vessel. The large solid organs remain out of reach for reasons that are mostly physical. That claim has a practical consequence. If you want to judge a tissue engineering result — a press release about a printed heart, a startup's pitch for a lab-grown liver, a paper reporting a new scaffold — the questions to ask are physical ones. How thick is it? How is oxygen reaching the middle? What is the flow, and what shear does it impose? What stiffness does the cell see? Was the material printed, and if so through what nozzle at what pressure? A reader who carries those questions will be harder to mislead and better able to recognize genuine progress when it comes. How the book proceeds The first chapter sets out the central problem: why living tissue cannot be made by simply putting cells in a mould, and why the diffusion distance of oxygen, a few hundred micrometres at most, governs nearly everything that follows. It also sketches the short history of the field, including its early, much-photographed demonstrations. The second chapter treats the scaffold as an engineered porous material. It works through porosity, pore size, interconnectivity, permeability and stiffness, and shows how these properties trade off against one another. The third and fourth chapters apply that framework to the most thoroughly studied case: hydroxyapatite and related calcium phosphates for bone. The third explains why bone mineral is the obvious model and what makes a synthetic ceramic osteoconductive or, more surprisingly, osteoinductive. The fourth is about fabrication: how a powder becomes a porous, load-bearing, resorbable part, by methods ranging from coating a polyurethane sponge to freezing a slurry to printing a ceramic paste line by line. The fifth chapter turns to the alternative of borrowing nature's own scaffold by stripping the cells out of a donor organ. Decellularization offers a matrix with the right architecture already built in, including the branching tree of the original vessels, and brings its own failure modes: residual DNA and immunogens, damaged basement membranes, and a recellularization problem far harder than the decellularization that precedes it. This chapter also records the field's most serious scandal, because it teaches something about how claims in this area should be evaluated. The sixth and seventh chapters concern bioreactors. The sixth explains why static culture fails in thick constructs and how stirred, rotating and perfusion systems change the transport picture, introducing the few dimensionless numbers that make the trade-offs legible. It takes up the question of shear stress in bone constructs, where flow is used both to feed cells and to signal to them. The seventh follows the same physics into the problem of vascularization: how endothelial cells read shear, how sprouting is promoted and suppressed by flow, and how engineers now build perfusable channels into constructs from the start. The eighth chapter is on bio-inks for 3D bioprinting, and it is mainly a chapter about rheology: viscosity, shear thinning, yield stress, recovery and crosslinking, and the narrow window in which a material is printable yet kind to the cells it carries. The ninth chapter is about the cells themselves — mesenchymal stromal cells, embryonic and induced pluripotent stem cells — and how the physical environment directs what they become. The tenth looks at translation: what has actually reached patients, what regulators require, and why manufacturing living products at scale is a problem in its own right. A conclusion draws the threads together and identifies what remains unsettled. What the reader needs No specialist background is assumed. Where an equation clarifies a point — Darcy's law for flow through a porous solid, the power-law model of a shear-thinning fluid, the parallel-plate formula for wall shear stress — it is introduced, explained in words, and used for something. The numbers that appear are ones that can be traced to the literature or derived from simple physics; where the literature disagrees, as it often does about optimal pore sizes or shear stress ranges, the disagreement is reported rather than smoothed over. Some vocabulary is unavoidable. A scaffold is any three-dimensional material on or in which cells are cultured or into which tissue grows. Extracellular matrix, or ECM, is the material cells secrete around themselves: collagens, proteoglycans, glycoproteins such as fibronectin and laminin, and, in bone, mineral. A construct is a scaffold together with whatever cells it carries. In vitro means in the laboratory; in vivo means in a living organism. A bioreactor is any device that controls the physical and chemical environment of a culturing construct. A bio-ink is a printable material that contains living cells. Other terms are defined where they first matter, and a short glossary at the end collects the ones that recur. One limitation should be stated plainly. This book is concerned with engineered tissues — structures built with scaffolds, bioreactors and printers — and touches only in passing on cell therapies that inject cells without a structure, on gene therapy, and on the large and important field of xenotransplantation, in which whole organs from genetically modified pigs are transplanted into people. Those fields overlap with tissue engineering and sometimes compete with it, and they are mentioned where the comparison is instructive. But the subject here is the engineering of the tissue itself: the material, the flow and the cells, and the physics that ties them together. Chapter 1. Why Tissue Is Not Cells in a Mould The simplest idea of how to make a piece of tissue is the one most people arrive at first. Take the cells the tissue is made of, grow a great many of them, and pack them into the shape you need. If a patient needs a new ear, grow cartilage cells and shape them like an ear. If a patient needs a liver, grow liver cells and pour them into a liver-shaped container. The idea fails for a reason that has nothing to do with biology in the narrow sense and everything to do with physics. Cells packed together at the density of a real tissue will consume the oxygen around them faster than oxygen can arrive. Within hours the cells at the centre are hypoxic; within a day or two they are dying. The block of tissue becomes a thin living shell around a necrotic core. Understanding why this happens, and how sharply it limits what can be built, is the foundation for everything else in this book. The arithmetic of oxygen Oxygen is poorly soluble in water. Culture medium in equilibrium with ordinary air at body temperature holds roughly 0.2 millimoles of dissolved oxygen per litre, something like a fortieth of what the same volume of arterial blood carries when its haemoglobin is saturated. In the body, haemoglobin solves the carriage problem and the capillary network solves the distance problem. In a dish or a scaffold, neither exists. The only way oxygen reaches a cell in the interior of a static construct is by diffusion: the random walk of molecules from regions of high concentration to regions of low concentration. Diffusion is efficient over short distances and hopeless over long ones, because the time it takes scales with the square of the distance. The diffusion coefficient of oxygen in water at 37 °C is roughly 3 × 10⁻⁹ square metres per second. A molecule will typically wander several tens of micrometres in a second, but to cover a millimetre it needs several minutes, and to cover a centimetre it needs several hours. While it wanders, cells along the way are taking oxygen out of the stream. There is a simple way to estimate how deep oxygen penetrates into a slab of tissue before it runs out. Suppose oxygen arrives at one face of the slab at concentration C₀, diffuses inward with diffusion coefficient D, and is consumed uniformly at a volumetric rate Q. Solving the steady diffusion equation with constant consumption gives a penetration depth L equal to the square root of 2DC₀ divided by Q. The formula is crude — real cells slow their consumption as oxygen falls, and real tissues are not uniform — but it gets the order of magnitude right, which is what matters. Now put in numbers. A typical mammalian cell in culture consumes oxygen at a rate somewhere between about 10⁻¹⁷ and a few times 10⁻¹⁶ moles per second; metabolically busy cells such as hepatocytes and cardiomyocytes sit at the upper end, and quiescent fibroblasts or chondrocytes at the lower. Real tissues contain somewhere between 10⁸ and 10⁹ cells per millilitre. Take a moderate consumption of 5 × 10⁻¹⁷ moles per cell per second and a cell density of 10⁸ per millilitre, and the penetration depth comes out at about half a millimetre. Raise the density to a tissue-like 10⁹ cells per millilitre and the depth falls to roughly 150 micrometres. Make the cells hepatocytes and it falls further. This is the origin of a number that appears throughout the tissue engineering literature: that cells in living tissue lie within about 100 to 200 micrometres of a capillary. The figure is not a biological accident. It is what diffusion permits. The body builds its vascular network densely enough that almost no metabolically active cell is ever farther from blood than oxygen can travel. August Krogh, the Danish physiologist who won the 1920 Nobel Prize for work on the regulation of capillaries, built the first quantitative model of this arrangement, treating each capillary as feeding a cylinder of surrounding tissue. The Krogh cylinder remains the standard first approximation, and its lesson is the same: tissue is organized around supply lines, and those supply lines are close together. For the engineer, the arithmetic draws a hard boundary. A static construct with tissue-like cell density can be no more than a few hundred micrometres thick if its cells are to survive. A slab of cartilage a millimetre or two thick can just about manage, because chondrocytes are sparse and consume little; cartilage in the body has no blood vessels at all and lives by diffusion from synovial fluid. A slab of heart muscle cannot, and a block of liver a centimetre on a side is out of the question. Everything bulky in tissue engineering — every bone graft larger than a few millimetres, every attempt at a solid organ — has to defeat this limit or work around it. Three ways around the limit There are only a few ways to beat diffusion, and the rest of this book is in large part an account of them. The first is to lower the demand. If the cells in a construct are few, or metabolically quiet, they can survive at greater depth. This is why many successful engineered tissues are seeded sparsely and allowed to mature slowly, and why a scaffold implanted with no cells at all can be any size: it asks the patient's own cells to migrate in from its surfaces, and they arrive gradually, with blood vessels following behind them. Most porous bone graft substitutes work this way. Their job is to be a structure into which bone and vessels can grow from the surrounding healthy tissue, and their size is limited not by the oxygen consumption of seeded cells but by how far the host's ingrowth will reach. The second is to improve the supply by moving fluid. If medium is pumped through a porous construct rather than left to sit around it, oxygen and nutrients are carried inward by flow, a process called convection, which is not limited by the square-of-distance penalty that cripples diffusion. Convection still has to deliver to within diffusion range of every cell, but the pores of the scaffold can be made small enough for that. This is the principle of the perfusion bioreactor, and its consequences, good and bad, occupy two later chapters. The catch is that moving fluid exerts a drag on the cells it passes, called shear stress, and cells respond to shear stress in strong and specific ways. The third is to build a vascular network into the construct, so that it can be connected to the patient's circulation, or to a bioreactor's circuit, and fed as a real organ is fed. This is the hardest of the three and the one on which the ambition to build solid organs depends. It can be pursued by printing channels, by stripping the cells from a donor organ and keeping its vascular tree, or by encouraging endothelial cells to organize themselves into vessels under the right chemical and mechanical cues. None of these approaches has yet produced an engineered solid organ that functions in a patient. Notice that each strategy is physical before it is biological. The choice of cells and growth factors matters, but it operates inside a set of constraints fixed by transport and mechanics. A brilliant choice of cell type in a construct that starves will produce a necrotic core of brilliantly chosen cells. Mechanics as the second constraint Oxygen is the most urgent constraint, but not the only one. Tissues are mechanical structures. Bone resists compression and bending; cartilage resists compression and shear while allowing nearly frictionless sliding; tendon resists tension; arteries resist pulsatile pressure; the heart generates force. An engineered replacement must survive the loads it will meet, at least until the patient's own tissue takes them over. This introduces a second set of trade-offs. A scaffold must be porous, because cells and vessels must get in and fluid must flow through, but porosity costs strength. A porous ceramic with 70 percent porosity has a small fraction of the strength of the dense material, and the relationship is steeply nonlinear, as later chapters explain. A soft hydrogel can be wonderfully hospitable to cells and almost useless as a structure. A material that degrades as new tissue forms — a goal of most scaffold designers, because nobody wants a permanent foreign body — must lose its strength on a schedule that matches the growth of strength in the tissue replacing it. Mechanics also enters at the scale of the single cell, and here the story becomes more interesting. Cells sense the stiffness of what they sit on, the forces transmitted to them through their attachments, and the drag of fluid flowing past them. They respond by changing shape, reorganizing their internal skeletons, moving, dividing, and switching on or off whole programmes of gene expression. Mechanical signals can decide whether a stem cell becomes bone or fat, whether an endothelial cell lines up with the flow or sprouts a new vessel, whether a chondrocyte makes cartilage matrix or fibrous scar. The physical design of a scaffold or a bioreactor is therefore not just about keeping cells alive. It is about telling them what to do. A short history of a young field People have repaired bodies with borrowed materials for a very long time. Surgeons were transplanting skin by the nineteenth century, and the twentieth century brought synthetic vascular grafts, joint replacements and the first successful kidney transplant, between identical twins in Boston in 1954. What was new in the late twentieth century was the idea of building living tissue deliberately, from cells and materials, as an engineering product. Several threads converged. Cell culture had matured; it was routine to grow many human cell types. Polymer chemists had developed biodegradable materials, especially the polyesters polylactic acid and polyglycolic acid, which were already used in absorbable sutures and could be processed into porous structures. Ceramicists had discovered that certain glasses and calcium phosphates bonded to living bone rather than being walled off by it; Larry Hench's bioactive glass, developed from the late 1960s, was the founding example. And there was demand: donor organs were scarce, and the shortage was worsening. In the late 1980s, meetings sponsored by the US National Science Foundation gave the field its name, and in 1993 Langer and Vacanti published their programmatic review in Science. They proposed that biodegradable polymer scaffolds seeded with cells could regenerate tissue, and they framed the field as an interdisciplinary enterprise combining engineering and life sciences. The review is widely cited as the field's founding statement. The field's most famous image arrived a few years later. In work published in 1997, a group including Charles Vacanti and Yilin Cao seeded bovine chondrocytes onto a biodegradable polymer scaffold in the shape of a human ear and implanted it under the skin of an immunodeficient mouse. Photographs of the hairless mouse with what looked like a human ear on its back circulated widely, and the image still turns up in discussions of the ethics of biotechnology, often with the mistaken belief that the mouse had been genetically engineered to grow an ear. In fact it was a demonstration that cartilage could form within a shaped scaffold, in a host whose immune system would not reject cells from another species. It is worth noticing which tissue was chosen: cartilage, which is avascular, sparsely cellular and nourished by diffusion. The demonstration was impressive precisely because it stayed within the physical limits described above. Progress in the following decades was real but uneven. Engineered skin substitutes reached the market in the late 1990s. A group led by Anthony Atala reported in The Lancet in 2006 on bladders built from patients' own urothelial and muscle cells on collagen and collagen–polyglycolic acid scaffolds and implanted in young patients with poorly functioning bladders. Engineered windpipes, urethras and blood vessels were implanted in small numbers of patients in the late 2000s and 2010s, with very mixed results, and one of these programmes ended in scandal. Laboratories learned to strip donor hearts, livers, lungs and kidneys of their cells, leaving ghostly translucent matrices that could be repopulated, at least partially. Three-dimensional bioprinting emerged as a technology in its own right, and with it a stream of press releases about printed organs. Through all this, a pattern held. The tissues that reached patients in significant numbers were thin, or avascular, or hollow with thin walls, or structurally forgiving. Large, thick, metabolically active, densely vascularized organs — the liver, kidney, heart and lungs that patients are actually waiting for — remained beyond reach. The reason, again and again, was the problem described at the start of this chapter. You cannot make a centimetre-thick block of living tissue without a way to feed its middle. What counts as success It is worth fixing early what a successful engineered tissue has to do, because the claims made for the field often blur the distinction between partial and full success. At a minimum, an engineered tissue has to survive implantation: its cells, if it has any, must remain alive long enough to integrate with the host, and the host must not reject or wall it off. It has to do a useful part of the native tissue's job: bear the load, line the lumen, secrete the protein, contract on command. It has to integrate: bone must bond to bone, vessels must connect to vessels, nerves must find their targets. And it has to last, either by being remodelled into native tissue or by remaining functional for the patient's lifetime. Many results that are reported as successes clear the first of these hurdles and not the later ones. A construct that is implanted and survives for a month in an animal, or that expresses tissue-specific genes in a dish, is a laboratory result, not a replacement organ. A large fraction of tissue engineering's real clinical impact has come not from complex living constructs but from relatively simple acellular materials — porous ceramics, processed animal tissues, collagen membranes — that provide a structure into which the body's own cells grow. That is not a failure of the field. It is evidence of where the physics allows success most easily. The chapters that follow begin there, with the scaffold as an engineered porous material, and with the most thoroughly studied application of that idea: synthetic bone. Chapter 2. The Scaffold as an Engineered Porous Material A scaffold is usually described by what it is made of: a ceramic, a polymer, a hydrogel, a decellularized matrix. That is the least useful way to describe it. Two scaffolds made of the same material can behave as differently as a sponge and a brick, while two made of different materials can serve the same purpose almost interchangeably. What decides a scaffold's fate is its architecture — how much empty space it contains, how that space is arranged, how big the openings are and how well they connect — together with the stiffness and degradation behaviour of the solid that frames the space. This chapter treats the scaffold as what it physically is: a porous solid through which cells must migrate, fluid must flow and load must pass. The jobs a scaffold must do It helps to list the jobs first, because they conflict. A scaffold must provide a surface to which cells can attach. Most tissue cells are anchorage dependent: deprived of attachment, they stop dividing and eventually die by a programmed route called anoikis. They attach through integrins, transmembrane receptors that bind specific short sequences in matrix proteins, the best known being the arginine–glycine–aspartate motif, abbreviated RGD, found in fibronectin and other adhesive proteins. Many synthetic materials have no such sequences of their own; cells stick to them only because proteins from the culture medium or from blood adsorb onto the surface within seconds, and the cells then bind to the adsorbed layer. The chemistry of the surface therefore matters indirectly, by deciding which proteins adsorb and in what conformation. A scaffold must allow cells and, in vivo, blood vessels to move through it. This requires openings large enough and connected enough for migration, which in practice means pores and the windows between them must be tens to hundreds of micrometres across. A scaffold must allow transport of oxygen, nutrients and wastes, by diffusion in static conditions and by convection in a perfusion bioreactor or after vascular ingrowth. That favours high porosity and good interconnectivity. A scaffold must carry load, at least enough to survive handling and implantation and, in a load-bearing site such as a long bone, enough to share the mechanical burden until new tissue forms. That favours low porosity. A scaffold should usually disappear, being degraded or resorbed as new tissue replaces it, on a timescale that matches tissue formation. Too fast, and the structure collapses before the tissue can bear load; too slow, and a foreign body persists, sometimes provoking chronic inflammation, sometimes simply blocking remodelling. And a scaffold must present the right cues: a surface chemistry, a stiffness, a topography and perhaps bound growth factors that tell cells what to become. There is no single optimum. Every design is a negotiated settlement among these demands, and the right settlement depends on the tissue, the site and whether the scaffold is to be seeded with cells or implanted bare. Porosity, pore size and interconnectivity Porosity is the fraction of a scaffold's volume that is empty. A scaffold with 70 percent porosity is 30 percent solid. Natural trabecular bone, the spongy bone inside vertebrae and at the ends of long bones, has a porosity that varies widely with site and age, commonly cited as somewhere between 50 and 90 percent. Cortical bone, the dense outer shell of long bones, has a porosity of only a few percent to around ten percent, made up of the canals that carry its blood vessels. Most bone scaffolds aim for something resembling trabecular bone, typically 50 to 90 percent porous. Porosity alone says nothing about the size of the pores, and pore size matters independently. The classic early experiments on bone ingrowth into porous ceramics, by Hulbert and colleagues around 1970, implanted calcium aluminate cylinders with different pore sizes in dogs and found that pores below roughly 100 micrometres admitted fibrous tissue and osteoid but not fully mineralized bone with its blood supply, while pores above about 100 micrometres admitted bone. That figure became a rule of thumb. Later work, summarized in an influential 2005 review by Karageorgiou and Kaplan, argued that larger pores, above about 300 micrometres, favour direct bone formation and the formation of capillaries, while smaller pores favour cartilage formation first, through hypoxia, before bone. Other studies have found good bone formation in a wide range of pore sizes, and the conflict is not fully resolved, partly because pore size is entangled with porosity, interconnectivity, surface area and stiffness in most experiments. The fair summary is that pores much below 100 micrometres are poor for bone ingrowth, that a range of a few hundred micrometres is generally favourable, and that very large pores sacrifice surface area and strength. There is also microporosity, a second, much finer level of porosity within the solid struts of a ceramic scaffold, with pores of a few micrometres or less. Microporosity arises naturally from incomplete sintering, and its importance was underappreciated for years. It increases surface area, raises the rate of dissolution and ion exchange, and allows proteins to adsorb in quantity. As the next chapter shows, it appears to be one of the features that lets some calcium phosphate ceramics induce bone in places where bone would not otherwise form. Interconnectivity is the extent to which pores are joined to one another, and the size of the openings, sometimes called windows or throats, between them. A foam with high porosity made of closed cells is almost useless as a scaffold: cells cannot pass between pores and fluid cannot flow. What matters for migration and transport is often the size of the narrowest window along a path, not the size of the pores themselves. A scaffold with 400-micrometre pores linked by 50-micrometre windows behaves in practice like a scaffold with small pores. Micro-computed tomography, which reconstructs the three-dimensional structure of a scaffold from X-ray projections, has become the standard tool for measuring interconnectivity and window size, replacing older methods that inferred structure from mercury intrusion or from two-dimensional sections. Permeability and Darcy's law The most useful single number for a scaffold through which fluid must flow is its permeability. Permeability is a property of the porous architecture, independent of the fluid, that describes how easily fluid passes through. It is defined by Darcy's law, which the French engineer Henry Darcy derived in 1856 from experiments on water flowing through sand filters for the fountains of Dijon. Darcy's law states that the flow rate per unit area through a porous medium equals the permeability divided by the fluid viscosity, multiplied by the pressure gradient driving the flow. Double the permeability and, at the same pressure, twice as much fluid flows. Permeability has units of area, usually square metres, and scaffold permeabilities span a huge range — many orders of magnitude, from very tight hydrogels to open printed lattices. It rises steeply with pore size and porosity. A rough theoretical relationship, the Kozeny–Carman equation, says permeability scales with the square of a characteristic pore or particle size and with a function of porosity that grows rapidly as porosity approaches one. The precise form matters less than the message: making pores twice as wide makes the scaffold roughly four times as permeable. Permeability matters in three ways. It determines how much pressure is needed to drive a given flow through the scaffold in a perfusion bioreactor, and therefore whether the scaffold will hold up or the seals will leak. It determines, together with the flow rate, the shear stress on cells lining the pores, which is the subject of a later chapter. And it appears to correlate with bone ingrowth in vivo, presumably because the same architecture that lets fluid through also lets cells and vessels through. Measuring permeability is straightforward: push fluid through a scaffold sample at known flow rate and measure the pressure drop, or the reverse. Mechanics of porous solids The mechanical cost of porosity is captured by a body of theory developed for cellular solids — foams, honeycombs, wood, bone — by Lorna Gibson and Michael Ashby, set out in their book Cellular Solids, first published in 1988. Their central result is that the properties of an open-cell foam depend on its relative density, which is the density of the foam divided by the density of the solid it is made from, or equivalently one minus the porosity. For open-cell foams that deform mainly by bending of their struts, the Gibson–Ashby analysis predicts that stiffness scales with the square of relative density, and that the crushing strength of a brittle foam scales roughly with relative density to the power of three halves. The consequence is dramatic. A ceramic scaffold at 70 percent porosity has a relative density of 0.3. Its stiffness should be about 0.09 of the dense material, or less than a tenth; its strength should be about 0.16 of the dense material, before considering any of the flaws that porous ceramics inevitably contain. At 85 percent porosity, the relative density is 0.15, stiffness falls to around 2 percent and strength to around 6 percent. Every gain in porosity costs disproportionately in mechanics. Architecture can soften this penalty. A structure whose struts carry load by stretching rather than bending — certain lattices, and ordered architectures designed by printing — can be much stiffer and stronger at the same density than a random foam. This is one of the main reasons engineers have been drawn to printed scaffolds: printing allows the arrangement of struts to be designed rather than accepted. Gradient architectures, dense at the edges and porous in the middle, imitate the transition from cortical to trabecular bone and can place strength where it is needed while keeping the interior open. Real bone sets the targets. Cortical bone has a compressive strength usually quoted in the range of roughly 100 to 200 megapascals and an elastic modulus in the range of roughly 10 to 20 gigapascals. Trabecular bone is far weaker and more variable, with compressive strengths of a few megapascals to perhaps 10 or more, depending on density and site. A porous ceramic scaffold can match trabecular bone's strength without much difficulty. Matching cortical bone in a porous structure has proved very hard, and no current porous ceramic scaffold is suitable for bearing full load in a large defect of a long bone on its own. Such defects are usually stabilized with metal plates, nails or frames while the scaffold and the new bone within it carry a growing share of the load. Brittleness is a further problem. Ceramics fail suddenly, by the propagation of cracks from flaws, and their strength varies widely from sample to sample. Engineers describe this variability with Weibull statistics, and ceramic scaffolds often have low Weibull moduli, meaning their strength is unpredictable. Toughness — the energy needed to break the material — matters as much as strength in a structure that will be screwed, cut and handled in an operating theatre. Natural bone is tough because it is a composite: stiff mineral crystals embedded in tough, flexible collagen. Many scaffold designs try to recapture that by combining a ceramic with a polymer. Degradation and its timing Most scaffolds are meant to be temporary. The ideal is often described as a scaffold that loses mass and strength at the same rate as new tissue gains it, so that the total load-bearing capacity of the site never dips dangerously. Different materials degrade by different mechanisms. The synthetic polyesters polylactic acid, polyglycolic acid and their copolymers degrade by hydrolysis of ester bonds, which proceeds throughout the bulk of the material rather than only at the surface. This bulk erosion has a disagreeable feature: the acidic degradation products accumulate inside the material and catalyse further hydrolysis, so degradation can accelerate suddenly, with a late burst of acid that may inflame surrounding tissue. Polycaprolactone, another polyester, degrades far more slowly, over a couple of years or more, which suits some applications and not others. Its low melting point, around 60 °C, makes it easy to print by melt extrusion, and it is widely used in printed bone scaffolds as a tough, slowly degrading framework. Calcium phosphate ceramics are resorbed partly by dissolution and partly by cells. Osteoclasts, the cells that break down bone, attach to the ceramic surface, seal off a zone beneath themselves and acidify it, dissolving mineral exactly as they would in natural bone remodelling. How quickly a ceramic dissolves depends strongly on its phase: as the next chapter explains, hydroxyapatite is much less soluble than tricalcium phosphate, so a scaffold's resorption rate can be tuned by mixing the two. Hydrogels — water-swollen polymer networks — degrade by hydrolysis, by enzymes, or by the reversal of whatever bonds hold them together. Natural hydrogels such as collagen, fibrin and hyaluronic acid are degraded by enzymes that cells themselves secrete, so cells can remodel them as they migrate. Synthetic hydrogels can be designed to do the same by including peptide sequences that are cleaved by the matrix metalloproteinases cells use to cut through matrix. The cell's view: stiffness and topography The properties discussed so far matter at the scale of the whole scaffold. Cells experience a different scale. A cell is a few tens of micrometres across. It sees the surface of a strut or the wall of a pore, and the stiffness that matters to it is the stiffness of the material immediately under its attachments. That stiffness has a surprisingly strong effect. Cells pull on their surroundings through their integrin attachments, using the contractile machinery of their internal skeleton. The resistance they meet shapes the tension inside the cell, and that tension feeds into signalling pathways that reach the nucleus. In a much-cited 2006 paper in Cell, Adam Engler, Dennis Discher and colleagues grew human mesenchymal stem cells on collagen-coated polyacrylamide gels of different stiffness. On the softest gels, with stiffness comparable to brain tissue, the cells took on neuron-like shapes and expressed neural markers; on gels of intermediate stiffness, comparable to muscle, they expressed muscle markers; on the stiffest gels, comparable to the collagenous precursor of bone, they expressed bone markers. Later work identified part of the mechanism in the transcriptional co-activators YAP and TAZ, which move into the nucleus when cells are on stiff substrates and spread out, and stay in the cytoplasm when cells are on soft ones. The finding has been refined and qualified a great deal since. Cells in three-dimensional gels respond differently from cells on flat gels; they respond to the rate at which their surroundings relax under stress, not only to stiffness; and what they actually sense may be the spacing of attachment points or the ability to remodel the matrix as much as the elastic modulus itself. But the central lesson has survived: a scaffold's material stiffness is an instruction, not merely a structural property. A rigid ceramic strut and a soft hydrogel present cells with very different messages, before any growth factor is added. Topography sends messages too. Cells align along grooves and fibres with dimensions comparable to or smaller than themselves, a phenomenon known since the early twentieth century as contact guidance. Electrospun fibres, produced by drawing a polymer solution into fine threads in a strong electric field, can be aligned to direct the growth of tendon, nerve or muscle cells. At the nanometre scale, surface roughness and pattern influence protein adsorption and cell adhesion in ways that are still being mapped. Material families compared Scaffold materials fall into a few broad families, each with characteristic strengths and weaknesses. Table 1 summarizes them; the rest of the book returns to each in more detail. Table 1. Principal families of scaffold material and their characteristic properties. Family Examples Mechanical character Degradation Typical use Calcium phosphate ceramics Hydroxyapatite, beta-tricalcium phosphate, biphasic mixtures Stiff, brittle, weak in tension Slow (HA) to moderate (TCP), cell- and dissolution-mediated Bone graft substitutes, coatings Bioactive glasses 45S5 Bioglass, borate glasses Stiff, brittle Dissolves, releasing ions Bone filler, dental, wound care Synthetic polyesters PLA, PGA, PLGA, polycaprolactone Tough, moderately stiff Bulk hydrolysis; weeks (PGA) to years (PCL) Printed frameworks, sutures, composites Natural hydrogels Collagen, gelatin, fibrin, alginate, hyaluronic acid Very soft Enzymatic or ionic; often fast Cell carriers, bio-inks, soft tissue Decellularized matrix Small intestinal submucosa, dermis, whole-organ matrix Native-like, tissue dependent Remodelled by host cells Soft tissue repair, organ scaffolds The categories blur at the edges. Composites of ceramic and polymer, hydrogels reinforced with printed polymer meshes, and decellularized matrix digested into a printable gel all combine the virtues of more than one family, and much current work is on such hybrids. The table is a map of the starting materials, not a menu of finished designs. The scaffold that does nothing There is one more point to make before turning to bone. Much of the clinical benefit of scaffolds has come from materials that do very little: no seeded cells, no growth factors, just a structure the body can grow into. This is sometimes called the conductive strategy, as opposed to an inductive one in which the scaffold actively tells cells what to become, or a cell-based one in which the scaffold arrives already populated. Conductive scaffolds dodge the diffusion limit, because they arrive with no cells that need oxygen. They dodge most of the regulatory difficulty of living products, because they are classed as devices rather than biological medicines. And they recruit the patient's own cells, which are perfectly matched and come with their own blood supply. Their limitation is reach: the body will grow into a scaffold only so far from healthy tissue, and in a very large defect, or in a patient whose healing capacity is impaired by age, radiation or disease, the ingrowth may never arrive. The more ambitious strategies in this book — seeded cells, bioreactor culture, prevascularized constructs — exist to extend that reach. But the conductive scaffold is the benchmark against which they must be judged, and in bone, where it has been studied longest, it is the place to begin. Chapter 3. Bone, Its Mineral, and the Case for Hydroxyapatite Bone is the tissue engineering has served best, and the reasons are instructive. Bone heals. A simple fracture, held still, will knit back together without scar, restoring the original tissue rather than patching it with something inferior — a capacity shared by few other tissues in the adult body. Bone is also, in large part, mineral, and the mineral is a calcium phosphate close in composition to compounds a chemist can make in a beaker. That combination, a tissue with a strong drive to regenerate and a major component that can be synthesized, made bone the natural proving ground for scaffolds. This chapter explains what bone is, what goes wrong when it cannot heal, and why hydroxyapatite and its relatives became the default materials for helping it. What bone is made of By weight, mature bone is roughly 60 to 70 percent mineral, 20 to 30 percent organic matter and around 10 percent water, with the proportions varying by bone, age and species. The organic matter is mostly type I collagen, arranged in fibrils that are themselves built from triple-helical molecules staggered in a regular pattern. The gaps in that pattern are where mineral first forms. The mineral is a calcium phosphate usually described as a poorly crystalline, carbonated apatite. Its ideal parent compound is hydroxyapatite, with the formula Ca₁₀(PO₄)₆(OH)₂ and a calcium-to-phosphorus molar ratio of 1.67. Biological apatite departs from the ideal in several ways. It contains several percent carbonate by weight, substituting for phosphate and hydroxyl groups; it contains small amounts of sodium, magnesium and other ions; it is deficient in calcium and hydroxyl; and its crystals are tiny, thin plates only a few nanometres thick and some tens of nanometres long. Those departures matter. The small size, the carbonate and the defects make biological apatite far more soluble and reactive than stoichiometric hydroxyapatite, which is exactly what the skeleton needs: a mineral reservoir that cells can dissolve and redeposit as the body regulates its blood calcium and repairs fatigue damage. The composite structure is the source of bone's mechanical character. Collagen alone is tough and flexible but not stiff; the mineral alone is stiff but brittle. Together, with mineral platelets embedded in and around collagen fibrils, they produce a material that is both stiff and tough, resisting crack propagation through a hierarchy of mechanisms operating from the nanometre scale of the fibril to the millimetre scale of the osteon. No synthetic material yet reproduces that combination fully. Bone is also intensely cellular and alive. Osteoblasts lay down new matrix and orchestrate its mineralization. Some of them become trapped in the matrix they make and turn into osteocytes, which sit in small cavities connected by a network of fine canals, and which are now understood to be the main mechanical sensors of bone, responding to the flow of fluid through their canal network when the bone is loaded. Osteoclasts, large multinucleated cells derived from the blood's monocyte lineage, resorb bone. The constant coupled activity of osteoclasts and osteoblasts, called remodelling, replaces old and damaged bone with new and adapts bone's architecture to the loads it carries. How bone heals, and when it does not When a bone breaks, the healing that follows is a sequence that resembles, in compressed form, the way the skeleton was first built in the embryo. Bleeding forms a haematoma around the break; inflammatory cells arrive and release signals; progenitor cells from the periosteum, the membrane covering the bone's surface, and from the marrow migrate in and multiply. Where the fracture is stable and well supplied with blood, they can form bone directly. Where there is movement or poor blood supply, they first form cartilage, which bridges the gap as a soft callus and is then replaced by bone, through a process that depends on blood vessels growing into the cartilage. Months later, remodelling restores the bone's original shape. Usually this works. Sometimes it does not. Fractures may fail to unite, especially when the blood supply is damaged, when the bone ends are unstable, when infection intervenes, or when the patient smokes, has diabetes or is old. And some gaps are simply too large. Above a certain size, a bone defect will not heal on its own however well it is stabilized. Such defects are called critical-size defects, and their existence is the reason bone grafting exists. They arise from high-energy trauma that destroys bone, from the removal of tumours, from infection that must be cleared by cutting out dead bone, and from the bone loss that accompanies failed joint replacements. In the jaw, the loss of bone after tooth extraction creates a smaller, commoner version of the same problem, and dental implantology is one of the largest markets for bone substitutes. There is no single critical size. It depends on the bone, the species, the age of the animal and the condition of the surrounding soft tissue. In animal research, standard defect models — a circular hole of fixed diameter in a rat's skull, a segment removed from a sheep's tibia — are used to test whether a scaffold can achieve what the body cannot. Their standardization makes results comparable, though how well a rodent skull predicts a human leg is always open to debate. The autograft standard and its costs The benchmark for treating a bone defect remains the autograft: bone taken from elsewhere in the same patient, most often the crest of the pelvis, and packed into the defect. Autograft has every property a scaffold designer would want. It is osteoconductive, providing a structure for new bone to grow along. It is osteoinductive, containing growth factors in its matrix that recruit and instruct progenitor cells. It is osteogenic, carrying living bone-forming cells. And it provokes no immune reaction. It also has serious drawbacks. The supply is limited by what can be taken from the patient's own skeleton. The harvest requires a second surgical site, which lengthens the operation and can leave lasting pain, nerve injury, infection or fracture at the donor site; the reported rates of such complications vary widely between studies. The quality of the graft depends on the health of the patient, so the people who most need help — the elderly, the ill, the irradiated — often have the poorest bone to give. Allograft, bone from deceased donors, avoids the second operation and is available in larger quantities, but processing to remove cells and reduce disease transmission strips most of its osteogenic and some of its osteoinductive capacity, and large structural allografts have a substantial rate of late complications, including fracture and failure to integrate. Demineralized bone matrix, allograft bone from which the mineral has been dissolved with acid to expose its growth factors, is osteoinductive to a variable degree that depends on the donor and the processing. Synthetic bone substitutes exist to escape these limits: unlimited supply, no second surgery, no disease transmission, and consistent quality. The price is that a synthetic material starts with none of autograft's biological activity, and every bit of it must be designed in. Three words: conduction, induction, osteogenesis The vocabulary just introduced is worth fixing, because it is used loosely in the literature and the distinctions matter. Osteoconduction is the property of a surface or structure that allows bone to grow on it and into it from adjacent bone. It is passive in the sense that the material provides a path rather than a command. An osteoconductive scaffold placed far from existing bone, under the skin or in muscle, will not form bone. Osteoinduction is the ability to cause undifferentiated progenitor cells to become bone-forming cells, and thus to form bone even in sites where bone does not normally form. The classic demonstration was by Marshall Urist, who reported in Science in 1965 that demineralized bone matrix implanted in muscle produced new bone. That observation led, over the following decades, to the identification of the bone morphogenetic proteins, or BMPs, a family of signalling molecules that can induce bone formation wherever they are placed in sufficient amounts. Osteogenesis, in this context, is the direct formation of bone by living cells delivered with the graft. Osseointegration is a related term, from the work of Per-Ingvar Brånemark on titanium dental implants, describing direct structural and functional contact between living bone and the surface of an implant, without an intervening layer of fibrous tissue. Why hydroxyapatite The logic of choosing hydroxyapatite as a bone substitute seems obvious: make the scaffold out of the same mineral that bone is made of. The history was less tidy. Calcium phosphates had been tried as bone fillers since the early twentieth century, but the modern interest dates to the 1970s and 1980s, when several groups showed that dense and porous hydroxyapatite implants were well tolerated and that bone grew directly onto their surfaces without an intervening fibrous capsule. That direct bonding is the property that defines a bioactive material, a term associated with Hench's work on bioactive glass, and it separates calcium phosphates and certain glasses from bioinert materials such as alumina, which bone tolerates but does not bond to. The mechanism of bonding is now reasonably well understood in outline. When a calcium phosphate ceramic is placed in a physiological fluid, its surface partly dissolves and exchanges ions with the surroundings, and a thin layer of carbonated apatite, similar to bone mineral, precipitates on it. Proteins from the surrounding fluid adsorb to this layer, cells attach to the proteins, and osteoblasts lay down collagen that mineralizes in continuity with the apatite surface. The ceramic becomes, in effect, part of the bone. The same apatite-forming tendency underlies a widely used laboratory test, in which a material is soaked in simulated body fluid, a solution with ion concentrations approximating human blood plasma, and inspected for apatite formation. The test, popularized by Tadashi Kokubo, is useful as a screen, though apatite formation in simulated fluid does not guarantee good performance in the body. The calcium phosphate family Hydroxyapatite is only one of several calcium phosphates used in bone repair, and the differences among them are mainly differences of solubility, which translate into differences of resorption rate. Table 2 compares the most important members of the family. Table 2. Calcium phosphates used in bone repair, compared by composition and behaviour. Compound Abbreviation Ca/P molar ratio Relative solubility at body pH Typical role Hydroxyapatite HA 1.67 Lowest of the group Slowly resorbed scaffolds, coatings on metal implants Beta-tricalcium phosphate β-TCP 1.50 Considerably higher than HA Resorbable granules and blocks Alpha-tricalcium phosphate α-TCP 1.50 Higher than β-TCP Reactive component of cements Biphasic calcium phosphate BCP Between 1.50 and 1.67 Tunable by HA:TCP ratio Scaffolds with tailored resorption Dicalcium phosphate dihydrate (brushite) DCPD 1.00 High Product of fast-setting cements Octacalcium phosphate OCP 1.33 Intermediate Proposed precursor of bone mineral The ranking of solubility in the table holds at physiological pH and reflects equilibrium chemistry; actual resorption rates in the body also depend on crystal size, microporosity, density and the activity of cells, so the table should be read as a guide to tendency, not a timetable. The practical consequence is that a designer can choose how long the scaffold should last. Pure, well-sintered hydroxyapatite resorbs extremely slowly; granules and blocks of it can persist in the body for years, sometimes indefinitely, embedded in new bone. That persistence is not always bad — the remaining ceramic is surrounded by bone and continues to bear load — but it frustrates the ideal of complete replacement by natural tissue and can interfere with later imaging or surgery. Beta-tricalcium phosphate resorbs much faster, often within months to a couple of years depending on its form, and sometimes faster than bone can replace it. Biphasic calcium phosphates, mixtures of the two, were developed in the 1980s, notably by Racquel LeGeros and Guy Daculsi and their colleagues, precisely to tune resorption between these extremes. A commonly used composition is around 60 percent hydroxyapatite and 40 percent beta-tricalcium phosphate, though many ratios have been produced and studied. Ion substitution offers another set of controls. Because the apatite lattice tolerates substitution, synthetic hydroxyapatite can be doped with carbonate to make it more like bone mineral and more soluble, with silicon, which has been reported to accelerate bone apposition, with strontium, which affects the balance of bone formation and resorption, and with magnesium, zinc and other ions. The evidence for a strong biological benefit of each substitution varies, and much of it comes from animal and cell culture studies; the general principle that the chemistry of the ceramic can be adjusted to modulate cellular responses is well established. Calcium phosphate cements A separate branch of the family deserves mention because it changes how the material reaches the defect. Calcium phosphate cements, developed from the 1980s onward with early work by LeRoy Chow and William Brown at the American Dental Association laboratories, are powders mixed with a liquid to form a paste that sets in the body at body temperature, through a dissolution and precipitation reaction that produces either apatite or brushite. The paste can be injected or moulded into irregular defects, filling them completely, which a pre-formed block cannot do. Cements trade architecture for convenience. A set cement is dense, with microporosity but usually without the large interconnected pores that allow bone and vessels to grow in, so resorption proceeds slowly from the surface inward. Researchers have introduced macroporosity by adding foaming agents, soluble particles that dissolve away, or degradable polymer fibres, which also increase toughness. Cements are also weak in tension and shear, so they are used mainly in sites that are protected from such loads, such as filling defects in the metaphysis of long bones beneath supported joint surfaces, or in augmenting screws in osteoporotic bone. Osteoinduction by ceramics The most surprising finding in the history of calcium phosphate scaffolds is that some of them are osteoinductive on their own. From the early 1990s onward, groups in the Netherlands, China and elsewhere — among them those of Klaas de Groot, Joost de Bruijn, Huipin Yuan and Xingdong Zhang — reported that certain porous calcium phosphate ceramics, implanted in the muscles of dogs, goats, pigs and baboons, formed bone with no added cells or growth factors. The effect is species dependent, appearing readily in larger animals and much less reliably in rats and mice, and it depends strongly on the ceramic's microstructure. Ceramics with high microporosity and large specific surface area, often sintered at lower temperatures, show it; dense, well-crystallized ceramics of the same chemistry generally do not. In a widely discussed comparison published in 2010 in the Proceedings of the National Academy of Sciences, Yuan, de Bruijn and colleagues compared two tricalcium phosphate ceramics of the same chemistry that differed in microstructure. The one with finer grains and more micropores formed bone in muscle and healed critical-size defects in sheep about as well as autograft in their model; the coarser one did not. The mechanism is still debated. Plausible contributions include the local release of calcium and phosphate ions, the adsorption and concentration of the body's own BMPs and other proteins on the high-surface-area ceramic, and the influence of surface topography on the behaviour of progenitor and immune cells, especially macrophages. Whatever the mechanism, the finding reinforced the main argument of this book: what the scaffold is physically, down to its micrometre-scale texture, tells cells what to do. Growth factors and their lesson If a scaffold cannot induce bone strongly enough by its structure, the obvious alternative is to load it with an inducing protein. Recombinant human BMP-2, delivered on an absorbable collagen sponge, was approved by the US Food and Drug Administration in 2002 for certain spinal fusion procedures, and later for some tibial fractures and dental applications. It is undeniably osteoinductive. Its subsequent history is a warning about the cost of forcing biology with high doses. Because a soluble protein diffuses away from a collagen sponge quickly, very large doses — milligrams, many orders of magnitude above the amounts present naturally — were used to ensure an effect. Concerns emerged about complications, including swelling in the neck after off-label use in anterior cervical spine surgery, about which the FDA issued a public health notification in 2008, as well as ectopic bone formation, inflammation and a debated association with cancer risk. An independent reanalysis of industry trial data, funded through Yale University and published in the Annals of Internal Medicine in 2013, concluded that the benefit over autograft in spinal fusion was modest and the evidence on harms incompletely reported in the original publications. The lesson for scaffold design is the one this book keeps returning to. A growth factor is only as good as the delivery system that holds it in place and releases it at the right rate. Much current work therefore aims to bind growth factors to scaffolds in ways that mimic how natural matrix holds them — through heparin, through engineered binding domains, through the affinity of the protein for mineral — so that smaller doses act locally and for longer. This is the scaffold designing the transport problem again, at the scale of a molecule instead of an oxygen supply. From the right material to the right structure Hydroxyapatite and its relatives provide the right chemistry for a bone scaffold. Chemistry alone does not make a scaffold. A dense block of the finest hydroxyapatite is a poor bone graft: bone will bond to its surface but cannot grow through it, and it will sit in the defect for years like a stone. What makes the material useful is the architecture into which it is formed — the porosity, the pore size, the interconnection, the microporosity of the struts, the strength of what remains. Turning a ceramic powder into such a structure, reproducibly and at a strength that survives surgery, is a problem of processing, and it is the subject of the next chapter. Hashtags: #TissueEngineeringAndScaffolding #TissueEngineering #ScaffoldDesign #Bioreactors #Decellularization #StemCells #ExtracellularMatrix #MassTransport #OxygenDiffusion #PerfusionBioreactors #FluidShearStress #Vascularization #PorousScaffolds #ScaffoldPermeability #DarcyLaw #Hydroxyapatite #CalciumPhosphateScaffolds #Osteoconduction #Osteoinduction #MesenchymalStemCells #InducedPluripotentStemCells #Bioinks #ThreeDimensionalBioprinting #Mechanotransduction #FutureOfRegenerativeMedicine

Download the Book (PDF): Introduction A woman in her late forties sits in a reclining chair in an outpatient clinic. A technician lowers a plastic-cased coil, shaped like a figure eight, against the left side of her forehead, a few centimetres above and in front of her ear. There is a series of sharp clicks, like a woodpecker tapping on a hollow branch, and with each burst she feels a firm tapping on her scalp and a faint twitch in the muscles around her eyebrow. Four seconds of clicking, twenty-six seconds of quiet, and then again. Thirty-seven minutes later she drives herself home. She will come back tomorrow, and the day after, five days a week for six weeks. She has tried six antidepressants, two augmentation strategies and two courses of psychotherapy over four years. Nothing has given her more than a few weeks of partial relief. Whether those six weeks help her depends on decisions most patients never see. Someone had to decide where on her head the coil should sit, and that decision rests on a chain of reasoning that runs from scalp landmarks to cortical anatomy to the large-scale networks of the brain. Someone had to decide how strong each pulse should be, and that decision rests on a measurement taken on her first day, when the coil was moved over the top of her head until a single pulse made her right thumb jump. Someone had to choose a pattern of pulses, a number of sessions, and a plan for what happens when the course ends. Each of these choices has a literature behind it, and each has been made badly in clinics and in trials often enough to change the results. This booklet is about those choices. Its subject is transcranial magnetic stimulation, usually shortened to TMS, as a treatment for depression that has not responded to adequate trials of medication and psychotherapy. Its argument is that TMS is best understood not as a gadget that either works or does not, but as a dosed intervention whose outcome is governed by two measurements and one plan. The two measurements are the target, meaning where in the brain the induced current is concentrated, and the intensity, meaning how much current is induced there relative to the individual's own cortical excitability. The plan is what happens after the acute course, because the question that matters most to patients with refractory illness is not whether they get better in six weeks but whether they stay better. When these three things are done well, TMS takes a defined and defensible place in the treatment of refractory depression. When they are done carelessly, the treatment looks weaker than it is, and comparisons with its older rival, electroconvulsive therapy, are distorted. Why this comparison matters Electroconvulsive therapy, or ECT, remains the most effective acute treatment for severe depression that psychiatry possesses. It induces a generalised seizure under general anaesthesia and muscle relaxation, usually two or three times a week for a course of six to twelve treatments. Its remission rates in severe, and particularly in psychotic, depression are higher than anything else available. It also carries the burdens of repeated anaesthesia, a period of confusion after each treatment, memory loss for events around the course that is occasionally lasting, and a stigma that keeps many patients who would benefit from ever considering it. TMS arrived in clinical psychiatry in the 1990s carrying an implicit promise: that it might do what ECT does without the seizure, the anaesthetic or the memory loss. That promise has been partly kept and partly broken. TMS is not as effective as ECT in the short term for the sickest patients. It is substantially better tolerated, it can be given to an awake patient who then goes to work, and in non-psychotic depression the gap in acute effectiveness is smaller than the reputation of either treatment suggests. And when one looks past the end of the acute course, both treatments share a problem that is underappreciated: a large share of patients who remit with either one relapse within a year unless something is done to prevent it. The comparison, then, is not a contest with a single winner. It is a question of sequence and fit. Which patient should be offered which treatment, in what order, and with what plan for maintenance? Answering that question requires understanding what each treatment does, how well it has been measured, and what the long-term numbers really show. How the argument unfolds The booklet begins by defining the problem it addresses. "Treatment-resistant depression" is a term used loosely, and the first chapter pins down what it means, how common it is, and why the probability of remission with each successive medication trial falls so steeply that non-pharmacological treatments become rational choices relatively early. The next chapter explains what TMS physically does: how a brief, intense magnetic field induces an electrical current in cortical tissue, why that current is confined to the outer few centimetres of the brain, and how repeated pulses at different frequencies are thought to strengthen or weaken synaptic connections. This is the foundation for everything that follows, because both targeting and dosing are attempts to control where and how strongly that current acts. Three chapters then deal with the two measurements at the heart of the argument. One explains why the left dorsolateral prefrontal cortex became the standard target, and how the understanding of that target has shifted from a patch of cortex to an entry point into a distributed network connected to the deep structures implicated in mood. A second chapter is about mapping the target in practice, comparing the crude ruler-based method that dominated early clinical work with scalp-based systems, MRI-guided neuronavigation, and connectivity-guided targeting derived from functional imaging. A third chapter is about the motor threshold: what it is, how it is measured, why it is used as a yardstick for prefrontal dosing even though it is measured over a different part of the brain, and what goes wrong when it is measured carelessly. With target and dose established, the booklet surveys the protocols in current use, from the conventional ten-pulse-per-second regime that underpinned the first regulatory approvals to theta burst stimulation, deep TMS, and the accelerated, connectivity-targeted schedule that compresses a course into five days. It then examines the evidence for acute efficacy, separating sham-controlled trials from the larger and more optimistic data from clinical registries. The final substantive chapters address the comparison with ECT. One examines head-to-head trials and meta-analyses of acute outcomes. The next turns to durability and relapse, which is where the comparison becomes most interesting and least settled. A last chapter brings safety, patient selection and treatment sequencing together into a practical account of where TMS belongs in the care of a patient with refractory depression. A note on evidence The literature on TMS is large, uneven and often funded by device manufacturers. Many trials are small. Sham conditions are imperfect, because patients can sometimes tell real stimulation from fake. Registries report better results than randomised trials, for reasons that are partly real and partly artefact. Head-to-head trials against ECT are few, mostly old, and conducted before modern targeting and dosing methods were available. Throughout, the booklet tries to say what is known with confidence, what is plausible but unproven, and what remains unknown. Where specific figures are given, they come from identifiable published studies, and the notes at the end list the main sources. The reader who finishes should be able to look at a TMS treatment plan and ask the right questions about it. Where is the coil going and why? How was the intensity set? What protocol is being used and what evidence supports it? What happens if the patient responds, and what happens if they relapse? And when ECT is on the table as an alternative, what does the evidence actually say about how the two compare, not just in the first month but over the year that follows? Chapter 1. The Problem of Refractory Depression Major depressive disorder is common, recurrent and, for a substantial minority of patients, stubborn. Most people who seek treatment for a depressive episode get better, either with the first medication they try, with a second or third, with psychotherapy, or with time. But a large group does not. They work through one antidepressant after another, often for years, accumulating side effects, disappointment and disability while the illness continues. This group is the population for whom transcranial magnetic stimulation and electroconvulsive therapy are most relevant, and it is worth being precise about who they are before discussing what can be done for them. What "treatment-resistant" means There is no single, universally agreed definition of treatment-resistant depression, and the lack of one has hampered research for decades. The most widely used working definition, adopted in regulatory guidance from the European Medicines Agency and in most contemporary trials, is failure to respond to at least two antidepressant treatments of adequate dose and duration in the current episode. Each element of that definition carries weight. "Failure to respond" usually means that symptoms did not fall by at least half on a standard rating scale. A patient who improves by 30 percent has not responded in this technical sense, though they may feel somewhat better. "Adequate dose" means a dose within the recognised therapeutic range, not a starting dose that was never increased because of side effects. "Adequate duration" is usually taken as at least four to six weeks at that dose, since most antidepressants take several weeks to show their effect. "In the current episode" means the failures must relate to the illness the patient has now, not a previous episode years ago that may have had a different character. These distinctions matter in practice because a large share of patients described as treatment-resistant turn out, on careful review, to have received inadequate trials. They took a low dose for three weeks, stopped because of nausea, took a different drug at a low dose, stopped again. Clinicians reviewing referrals for TMS or ECT routinely find this pattern. The term for it is pseudo-resistance, and it is not a trivial matter: a patient with pseudo-resistance may do well with a properly conducted medication trial and may not need neuromodulation at all. Equally, a patient's history may contain undiagnosed bipolar disorder, an untreated thyroid disorder, ongoing alcohol misuse, or a medication that worsens mood, any of which can make depression look resistant to treatment when the real problem lies elsewhere. Several formal staging systems have tried to capture resistance as a continuum rather than a yes-or-no category. The Thase and Rush model, proposed in 1997, places patients in stages according to the number and class of failed treatments, with failure of a tricyclic antidepressant or a monoamine oxidase inhibitor counting as more severe resistance than failure of a newer drug, and failure of ECT marking the highest stage. The Maudsley Staging Method, developed by Fekadu and colleagues in 2009, adds the duration of the episode and its severity to the count of failed treatments, and also gives weight to whether augmentation and ECT have been tried. The Maudsley method has the advantage of predicting outcomes reasonably well: patients with higher scores are less likely to recover over subsequent years. For the purposes of this booklet, the distinction that matters most is between patients who have failed two or three medication trials and patients who have failed many treatments, including augmentation and perhaps ECT. These are different populations with different prospects, and much confusion in the TMS literature comes from lumping them together. The arithmetic of sequential failure The strongest single source of evidence about what happens when depression does not respond to the first treatment is the Sequenced Treatment Alternatives to Relieve Depression study, known as STAR*D. Funded by the United States National Institute of Mental Health and reported mainly in 2006, it enrolled several thousand outpatients with non-psychotic major depression in primary care and psychiatric clinics and treated them in a series of steps. All began with citalopram. Those who did not remit could move to a second step, which offered switching to another drug, augmenting with another drug, or cognitive therapy. Those who still did not remit moved to a third step, and then a fourth. The headline finding, reported by Rush and colleagues in the American Journal of Psychiatry, was that remission rates fell sharply at each step. About 37 percent of patients remitted at the first step, about 31 percent of those who entered the second step remitted there, but only about 14 percent at the third step and 13 percent at the fourth. Relapse also became more common the more steps a patient had needed before remitting. A patient who remitted only after three or four treatments was considerably more likely to relapse during follow-up than one who remitted on the first. The STAR*D investigators also reported a theoretical cumulative remission rate of about 67 percent across all four steps, a figure that was widely quoted as evidence that persistence pays. That figure has since been challenged. A reanalysis published in 2023 by Pigott and colleagues, which applied the study's original protocol-specified outcome measure and excluded patients who did not meet entry criteria, arrived at a much lower cumulative remission rate. The details of that dispute need not concern us here, but the direction of the correction is important: if anything, STAR*D overstated how well patients do when they keep trying one medication after another. The shape of the curve is what matters for treatment planning. After the second failed medication trial, the chance that the next medication will produce remission is low, on the order of one in seven or one in eight, and the chance that any remission obtained will last is lower still. This is the arithmetic that makes non-pharmacological treatments rational choices at a relatively early stage. If a third antidepressant offers a roughly 14 percent chance of remission, then a treatment offering a 30 percent or 50 percent chance deserves serious consideration, even if it is more burdensome to deliver. The human and economic weight of persistent depression Treatment-resistant depression is not simply depression that lasts longer. Patients with resistant illness have higher rates of suicide attempts and completed suicide, more physical illness, more hospital admissions and greater loss of work. Studies of health service use have consistently found that patients with resistant depression cost health systems considerably more than patients whose depression responds to first-line treatment, and that the indirect costs from lost employment and caregiving exceed the direct medical costs. Estimates vary with definitions and settings, but the pattern is consistent across countries. There is also a cost that rating scales capture poorly: the cumulative effect of years spent unwell. A patient who spends four years cycling through medications has lost four years of work, relationships and ordinary pleasures. Some will have lost a job, a marriage or the capacity to look after children. Many will have developed a firm belief that nothing will help them. That belief, which clinicians sometimes call therapeutic nihilism, can itself become an obstacle, because it makes patients reluctant to try something new and quick to abandon it if early sessions bring no change. The chronicity of the illness also matters biologically. Longer episodes are associated with lower response rates to most treatments, including both TMS and ECT. In TMS studies, the duration of the current episode and the number of failed medication trials are among the most consistent predictors of poorer outcome. A patient who receives TMS after two failed trials in an eighteen-month episode has a better chance than one who receives it after eight failed trials in a six-year episode. This is one reason the argument of this booklet places weight on timing: the same treatment, delivered at a different point in the illness, can produce markedly different results. The landscape of options after medication fails When two or more adequate medication trials have failed, a clinician has several broad paths. Each has a different evidence base, a different profile of burdens, and a different relationship to TMS. The first path is to continue with pharmacology using more aggressive strategies. These include augmentation with lithium, which has the strongest evidence of any augmenting agent and a specific anti-suicidal effect; augmentation with an atypical antipsychotic such as aripiprazole, quetiapine or brexpiprazole, several of which have regulatory approval for this use; switching to a tricyclic antidepressant or a monoamine oxidase inhibitor; or combining two antidepressants with complementary mechanisms. These strategies help some patients, but each also adds side effects, and augmentation with antipsychotics carries metabolic and movement-disorder risks that accumulate over time. The second path involves the rapid-acting glutamatergic agents. Intravenous ketamine, used off-label, and its enantiomer esketamine, approved as a nasal spray in the United States and Europe in 2019 for treatment-resistant depression, can produce improvement within hours to days. Their effects are often transient unless the drug is repeated, and they require supervised administration. In 2023 a large randomised trial known as ELEKT-D, reported by Anand and colleagues in the New England Journal of Medicine, compared intravenous ketamine with ECT in 403 patients with non-psychotic treatment-resistant depression and found ketamine non-inferior, with response rates of about 55 percent for ketamine against about 41 percent for ECT. That trial has been debated, particularly regarding how ECT was delivered and the high proportion of outpatients enrolled, but it has reshaped the discussion of what comes after medication failure. The third path is psychotherapy, particularly structured approaches such as cognitive behavioural therapy and its variants for chronic depression. Psychotherapy combined with medication tends to outperform either alone in chronic depression, though access is often limited and patients with severe symptoms may find it difficult to engage. The fourth path is neuromodulation. This includes ECT, the oldest and most effective; TMS in its various forms; vagus nerve stimulation, an implanted device approved for chronic resistant depression in the United States since 2005, with slow and modest but sometimes durable effects; deep brain stimulation, still experimental for depression; and transcranial direct current stimulation, a weak, cheap, portable form of electrical stimulation whose efficacy in depression is modest and debated. Magnetic seizure therapy, which uses a magnetic coil to induce a seizure under anaesthesia, sits between TMS and ECT and remains investigational. These paths are not mutually exclusive. Many patients receive TMS while continuing medication. Some receive ketamine and then TMS, or TMS and then ECT. The question is not which single treatment is best, but which sequence gives a particular patient the best chance of recovery with the least harm. Where TMS entered the picture TMS was first approved for depression in the United States in October 2008, when the Food and Drug Administration cleared the NeuroStar system for adults with major depressive disorder who had failed to achieve satisfactory improvement from one prior antidepressant at or above the minimal effective dose and duration in the current episode. That indication was notably narrow: one failed trial, not two. It reflected the design of the pivotal trial, which had found a clearer benefit in patients with less resistant illness. Over the following years, further devices were cleared, insurers developed coverage policies that typically required failure of several medication trials, and the treatment spread through outpatient psychiatry in North America. Canada, Australia, the United Kingdom and much of Europe adopted it in parallel, with professional bodies and health technology assessment agencies issuing their own guidance. In the United Kingdom, the National Institute for Health and Care Excellence concluded in 2015 that the evidence on efficacy and safety was adequate to support its use under normal arrangements for clinical governance. From the beginning, TMS occupied an awkward position. It was more burdensome than taking a pill, requiring daily attendance for several weeks. It was less effective than ECT in the sickest patients. Its early trials produced modest effect sizes, and some were negative. Sceptics argued that its benefits might be largely placebo, given the ritual and attention involved in daily treatment. Enthusiasts pointed to real responders, some with long histories of resistance, who recovered after TMS when nothing else had worked. Much of what follows in this booklet is an attempt to explain that ambivalence. A significant part of the gap between the sceptics and the enthusiasts, the argument goes, came from variation in how TMS was delivered: where the coil was placed, how intensity was set, how many pulses were given, and what was done after the course ended. As those parameters have been refined, the treatment has become more effective and more predictable, though not yet as predictable as anyone would like. Defining success A last preliminary matter is what counts as success. Clinical trials in depression usually report two outcomes. Response means a reduction of at least 50 percent in the score on a depression rating scale. Remission means a score below a threshold that indicates few or no symptoms: for example, 7 or less on the 17-item Hamilton Rating Scale for Depression, 10 or less on the Montgomery–Åsberg Depression Rating Scale, or 4 or less on the self-rated nine-item Patient Health Questionnaire. Some studies use other thresholds, which complicates comparison. Remission is the outcome that matters most for long-term prognosis. Patients who respond but do not remit, who are better but still have residual symptoms, are considerably more likely to relapse than those who remit fully. In treatment-resistant depression, response without remission is common, and it is worth noting at the outset that response rates for any treatment will always look more impressive than remission rates. Beyond remission, the outcomes that matter to patients include whether improvement lasts, whether they can return to work and relationships, whether side effects are tolerable, and whether the treatment can be repeated if the illness returns. These longer-term outcomes are where the comparison between TMS and ECT is least well studied and most consequential. A treatment that produces remission in half of patients but loses most of them within six months may be less valuable than one that produces remission in a third but keeps most of them well for a year. Holding both numbers in mind, the acute rate and the durability, is the discipline this booklet asks of its reader. The remaining chapters begin with the physical basis of TMS, because neither targeting nor dosing makes sense without an understanding of what a magnetic pulse actually does to the brain. Chapter 2. What a Magnetic Pulse Does to the Brain The name "transcranial magnetic stimulation" is slightly misleading. The magnetic field is the means, not the end. What actually stimulates neurons is an electrical current, induced in the brain by a rapidly changing magnetic field, in the same way that a changing magnetic field induces a current in the coils of an electrical generator. The skull, which is a poor conductor of electricity and resists the passage of current applied from outside, is almost transparent to magnetic fields. This is the central trick of TMS: it delivers electrical stimulation to the cortex without passing current through the scalp and skull, and therefore without the pain and the diffuse spread that made earlier attempts at transcranial electrical stimulation in conscious people impractical. From capacitor to cortex A TMS device consists of a bank of capacitors, a switch, and a coil of copper wire encased in plastic. The capacitors are charged to a high voltage, typically in the range of one to three thousand volts. When the switch closes, the capacitors discharge through the coil in a fraction of a millisecond, producing a current that peaks at several thousand amperes. That current generates a magnetic field perpendicular to the plane of the coil, reaching a strength of one to two and a half tesla at the coil surface, comparable to the static field inside a clinical MRI scanner. The crucial difference is that the TMS field rises and falls within a few hundred microseconds. According to Faraday's law of induction, a magnetic field that changes rapidly induces an electric field in any conductor nearby. Brain tissue, bathed in salty fluid, is a conductor. The induced electric field drives a current through the tissue in a loop roughly parallel to the coil but flowing in the opposite direction to the current in the coil. Where that induced current is strong enough and oriented appropriately along the long axis of a nerve fibre, it depolarises the membrane past its threshold and triggers an action potential. The neurons most readily excited are not the large cell bodies in the deep layers of the cortex but the axons, especially where they bend or terminate, because these are the points where the induced field produces the largest change in membrane voltage. In practice, a single TMS pulse over the motor cortex activates mainly interneurons and axons running horizontally in the superficial layers, which in turn excite the large pyramidal cells that project down the spinal cord. This is why a pulse over the motor cortex produces a series of waves of descending activity rather than a single volley, and why the orientation of the coil matters: rotating it changes which population of fibres is preferentially activated. The depth problem The induced electric field falls off steeply with distance from the coil. This is the single most important physical constraint on TMS. A standard figure-of-eight coil, with two circular windings placed side by side so that the currents beneath their junction add together, produces a focal field that is strongest directly under the centre of the coil and effective to a depth of about 1.5 to 2.5 centimetres below the scalp. Since the scalp and skull together are roughly 1 to 1.5 centimetres thick, this means a figure-of-eight coil stimulates only the crowns of the cortical gyri nearest the surface. It does not directly reach structures buried in the depths of sulci, let alone deep structures such as the subgenual anterior cingulate cortex, the amygdala or the nucleus accumbens. There is a fundamental trade-off between depth and focality. Coils can be designed to push the field deeper, but only by spreading it over a wider area of the surface. The H-coil used in what is marketed as deep TMS is a complex winding designed to produce a field that decays more slowly with depth, reaching perhaps 3 to 5 centimetres, at the cost of stimulating a broad swathe of prefrontal cortex rather than a focused spot. No coil design can produce a field that is stronger at depth than at the surface, because of the physics of how fields propagate in a conducting medium. Deep structures can be reached only through stronger surface stimulation, which increases discomfort and seizure risk, or through their connections with superficial regions. This last point is the key to understanding why TMS works for depression at all. The brain regions most consistently implicated in depressed mood lie deep: the subgenual cingulate, the amygdala, the ventral striatum. TMS cannot reach them directly. What it can do is stimulate a superficial region that is connected to them, and rely on the stimulation propagating along those connections. The choice of target, the subject of the next three chapters, is therefore a choice about which superficial entry point gives the best access to the deep network that matters. The distance from coil to cortex varies between individuals. Older patients, whose brains have shrunk with age and whose cerebrospinal fluid spaces have widened, may have prefrontal cortex lying farther from the scalp than younger patients. Because the field falls off steeply, a few millimetres of extra distance can substantially reduce the strength of the field reaching the cortex. This has direct consequences for dosing, discussed in the chapter on motor threshold. Single pulses, repetitive trains and lasting change A single TMS pulse produces a brief burst of neural firing that lasts a few tens of milliseconds and then fades. Over the motor cortex, this is enough to make a muscle twitch. Over the visual cortex, it can produce a flash of light. Over the prefrontal cortex, it produces nothing the person notices. Single pulses are useful for research and for measuring cortical excitability, but they do not produce lasting changes in mood. For therapeutic effects, pulses must be delivered repeatedly, in trains, over many minutes, on many days. This is repetitive TMS, or rTMS. The working hypothesis for how repeated stimulation produces lasting effects draws on the large body of research on synaptic plasticity, particularly long-term potentiation and long-term depression. In animal brain tissue, stimulating a pathway at high frequency strengthens the synapses along it, an effect that can last hours or longer, while prolonged low-frequency stimulation weakens them. Both processes depend on the NMDA type of glutamate receptor and on intracellular calcium signalling, and both are considered cellular mechanisms underlying learning and memory. Early human studies found that rTMS over the motor cortex at frequencies of around 5 hertz or higher tended to increase the size of the muscle responses to subsequent single pulses for some minutes afterwards, while stimulation at 1 hertz tended to decrease them. By analogy with the animal work, high-frequency rTMS came to be described as excitatory and low-frequency rTMS as inhibitory. This framework shaped the design of depression treatments: high-frequency stimulation was applied to the left dorsolateral prefrontal cortex, which imaging studies had suggested was underactive in depression, and low-frequency stimulation to the right side, which some studies suggested was relatively overactive. The framework is useful but oversimplified. The effects of a given frequency vary considerably from person to person; a substantial minority of healthy volunteers show the opposite of the expected effect, or none at all. The effects also depend on the brain's state at the time of stimulation, on prior activity, on the intensity of stimulation and on the number of pulses. Applying more stimulation does not always produce more effect, and in some studies doubling the duration of a protocol reversed its direction. The excitatory-inhibitory labels should be read as rough tendencies rather than reliable rules. Theta burst stimulation, which has become central to modern depression treatment, was derived directly from the plasticity literature. In the hippocampus, neurons naturally fire in bursts at around 5 hertz, the theta rhythm, and stimulating in this pattern is a particularly efficient way to induce long-term potentiation. In 2005 Huang and colleagues, working in London, adapted the pattern for human TMS: bursts of three pulses at 50 hertz, repeated five times per second. Delivered in an intermittent pattern, with two seconds of bursts followed by eight seconds of rest, it increased motor cortex excitability for up to an hour after only about three minutes of stimulation. Delivered continuously for forty seconds, it decreased excitability. Intermittent theta burst stimulation, abbreviated iTBS, became the basis of the shortest and most widely adopted modern protocols. From local change to network change If rTMS simply strengthened synapses in the patch of cortex under the coil, its antidepressant effects would be hard to explain. The prefrontal cortex is not the seat of mood; it is one node in a widely distributed set of interconnected regions. The modern understanding is that TMS changes the activity of the networks to which the stimulated region belongs, and that its therapeutic effect depends on which network is engaged. Several kinds of evidence support this. Studies combining TMS with functional imaging have shown that stimulating the prefrontal cortex changes activity not only beneath the coil but in distant regions connected to it, including the cingulate cortex, the thalamus and the striatum. Studies of patients before and after a course of rTMS have reported changes in the strength of functional connections between prefrontal regions and the subgenual cingulate, and some have found that the degree of change correlates with clinical improvement. And, as the next chapter discusses in detail, retrospective analyses of where the coil was placed have found that patients whose stimulation site was more strongly anticorrelated with the subgenual cingulate tended to improve more. The network view has an important practical consequence. If what matters is the connection between the stimulation site and a deep region, then the best site may differ from person to person, because the precise pattern of connectivity differs between individuals. This insight is the basis of connectivity-guided targeting, the most significant development in the field in the past decade. Other biological effects Beyond synaptic plasticity and network effects, rTMS has been reported to produce a range of other biological changes in animal and human studies: increased levels of brain-derived neurotrophic factor, a protein that supports neuronal growth and is reduced in some depressed patients; changes in monoamine release, particularly dopamine in the striatum; changes in the balance of excitatory and inhibitory neurotransmission measured with magnetic resonance spectroscopy; normalisation of the stress hormone system; and changes in gene expression. Many of these findings come from small studies, and few have been consistently replicated in humans with depression. They suggest that rTMS engages multiple biological systems, as one would expect of any treatment that alters brain activity repeatedly over weeks. None is yet established as the mechanism of its antidepressant effect. How ECT differs The contrast with ECT helps clarify what TMS does and does not do. ECT passes an electrical current directly through the scalp and skull, using electrodes placed either on both temples, on the right temple and the top of the head, or on both sides of the forehead. Because the skull resists the current, most of it is shunted around the scalp, and the stimulus must be strong enough to push enough current through the skull to trigger a generalised seizure. That seizure, spreading through both hemispheres and engaging deep structures, is the therapeutic event. The quality of the seizure, its duration, its spread and the way it terminates, are related to the strength of the antidepressant effect. TMS is designed to avoid a seizure. It stimulates a localised region of superficial cortex at intensities well below those that would trigger generalised activity, and relies on repeated stimulation to produce cumulative change. ECT acts on the entire brain at once, powerfully and briefly; TMS acts on a small region, gently and repeatedly. This difference accounts for much of what distinguishes them clinically: the greater acute potency of ECT, its cognitive side effects, the need for anaesthesia, and the greater precision demanded of TMS. ECT does not need to be aimed precisely, because the seizure generalises regardless of where it starts, though electrode placement does affect cognitive side effects. TMS must be aimed precisely, because it acts only where it is aimed and through the connections of that site. Magnetic seizure therapy combines features of both, using a high-powered magnetic coil to induce a seizure under anaesthesia. Because the magnetic field induces current directly in the cortex without passing through the scalp, the seizure can be initiated in a more focal region, and early studies suggest fewer cognitive side effects than ECT. It is not yet in routine clinical use. What this means for treatment design The physics and physiology of TMS lead to three practical principles. First, because the induced field falls off steeply with depth and is concentrated under the coil, the precise placement of the coil determines which cortical tissue is stimulated, and therefore which networks are engaged. Second, because the threshold for activating neurons depends on the individual's anatomy and excitability, the intensity of stimulation must be calibrated to each person rather than set at a fixed value. Third, because therapeutic effects accumulate through repeated sessions and depend on plasticity processes that vary in duration, the schedule and total amount of stimulation matter, as does the question of what happens when stimulation stops. These three principles correspond to the three elements of the booklet's argument: target, dose and durability. The next chapter takes up the first of them, by asking why the dorsolateral prefrontal cortex became the target of choice and what that target actually represents. Chapter 3. Why the Dorsolateral Prefrontal Cortex Almost every TMS treatment for depression delivered anywhere in the world today begins with the coil over the left dorsolateral prefrontal cortex. This choice is so standard that it is easy to forget it was ever a choice. It rests on a chain of observations stretching back to the 1980s, and the reasons it works, to the extent they are understood, are different from the reasons it was first chosen. Understanding both the history and the current rationale explains why targeting has become the most active area of research in the field, and why getting it right matters so much. What the dorsolateral prefrontal cortex is The prefrontal cortex is the front part of the frontal lobe, lying ahead of the motor and premotor areas. It is the region of the brain most expanded in humans compared with other primates, and it is involved in planning, working memory, flexible attention, the regulation of emotion and the control of behaviour in pursuit of goals. It is conventionally divided into several subregions according to their position and connections. The dorsolateral part lies on the upper outer surface, roughly beneath the area of the forehead above the outer end of the eyebrow and back towards the hairline. In the classical cytoarchitectonic map drawn by Korbinian Brodmann in the early twentieth century, the dorsolateral prefrontal cortex corresponds mainly to areas 9 and 46, with parts of areas 8 and 10. Functionally, the dorsolateral prefrontal cortex is a hub of what neuroscientists call the central executive network, or frontoparietal control network: a set of regions, including parts of the parietal cortex, that become active when a person holds information in mind, switches between tasks, or deliberately directs attention. It is well connected to other cortical regions and, through indirect pathways, to the cingulate cortex, the striatum and the thalamus. It is not a single homogeneous area. Different parts of it connect to different networks, and the boundaries between subregions vary from person to person. Its size is relevant to targeting. The dorsolateral prefrontal cortex as a whole covers several square centimetres of cortical surface on each side. A figure-of-eight coil stimulates effectively an area of roughly one to a few square centimetres, depending on intensity. So "targeting the dorsolateral prefrontal cortex" is not precise. The coil can be over the front, back, top or bottom of the region, and each placement stimulates a different patch with different connections. The original rationale The choice of the left dorsolateral prefrontal cortex grew out of three strands of evidence available in the early 1990s. The first came from neurology. Patients with strokes affecting the left frontal lobe seemed more likely to become depressed than patients with right frontal strokes, an observation reported by Robert Robinson and colleagues in the 1980s. The finding was later disputed, and subsequent meta-analyses did not consistently confirm a left-sided predominance, but at the time it suggested that loss of left frontal function might contribute to depression. The second came from functional imaging. Early studies using positron emission tomography and single photon emission tomography reported reduced blood flow and glucose metabolism in the left prefrontal cortex in depressed patients, with some normalisation after successful treatment. Like the stroke findings, these results were not perfectly consistent, but they pointed in the same direction: the left prefrontal cortex appeared underactive in depression. The third came from ECT. Studies by Harold Sackeim and others had shown that the antidepressant effect of ECT was related to the intensity of the stimulus delivered to the prefrontal regions, and that ECT reduced prefrontal activity in ways related to clinical outcome. This led to the hypothesis that the prefrontal cortex was a key site of action for ECT, and that stimulating it directly might reproduce some of ECT's benefits. Putting these together, Mark George, then at the National Institutes of Health, and colleagues reasoned that high-frequency rTMS, thought to be excitatory, applied to the left dorsolateral prefrontal cortex, thought to be underactive, might relieve depression. Their first open study, published in 1995, involved six patients with highly resistant depression; two showed substantial improvement. Alvaro Pascual-Leone and colleagues in Spain published a sham-controlled crossover study in 1996 in seventeen patients with psychotic depression, reporting marked improvement after left dorsolateral stimulation but not after stimulation at other sites or sham. These early studies were small, but they established the left dorsolateral prefrontal cortex as the default target, and nearly all subsequent clinical trials followed suit. The low-frequency right-sided approach developed in parallel. On the hypothesis that depression involves an imbalance between hypoactive left and relatively hyperactive right prefrontal regions, some groups applied 1 hertz stimulation, thought to be inhibitory, to the right dorsolateral prefrontal cortex. Trials of this approach have produced effects broadly comparable to left-sided high-frequency stimulation, and it remains a useful option, particularly for patients at elevated seizure risk or those who find high-frequency stimulation uncomfortable. Bilateral protocols, combining both, have also been tested, without clear evidence that they outperform either side alone. From region to network Over the following two decades, the rationale shifted. The simple model of a hypoactive left prefrontal cortex in need of activation gave way to a network model of depression, informed by functional imaging studies of intrinsic brain activity and by an unexpected development in neurosurgery. In 2005 Helen Mayberg and colleagues reported in Neuron that chronic electrical stimulation of white matter adjacent to the subgenual anterior cingulate cortex, a small region beneath the front of the corpus callosum also known as Brodmann area 25, produced marked improvement in several patients with severely resistant depression. Mayberg's earlier imaging work had shown that this region was overactive in depression and that its activity decreased with successful treatment by a range of methods, including medication, psychotherapy and ECT. Its activity seemed to be a common pathway of recovery. The subgenual cingulate is connected to the hypothalamus and brainstem, which regulate sleep, appetite and stress hormones; to the amygdala and insula, which are involved in emotional responses and bodily sensation; to the orbitofrontal cortex; and indirectly to the dorsolateral prefrontal cortex. That last connection is the important one for TMS. Studies of resting-state functional connectivity, which measure how activity in different brain regions rises and falls together while a person lies quietly in a scanner, revealed that the subgenual cingulate and parts of the dorsolateral prefrontal cortex tend to be anticorrelated: when one is more active, the other tends to be less active. This suggested a mechanism for TMS. Stimulating the part of the dorsolateral prefrontal cortex most strongly anticorrelated with the subgenual cingulate might, through network interactions, reduce subgenual activity in the same way that successful antidepressant treatments do. The Fox analysis The paper that crystallised this idea was published in 2012 by Michael Fox and colleagues, working at Harvard, in Biological Psychiatry. Its design was elegant. The authors collected the coordinates of stimulation sites from published TMS studies in depression, including the different targeting methods those studies had used, and examined each site's resting-state connectivity with the subgenual cingulate using brain scans from a large sample of healthy people. They found that sites associated with better clinical outcomes tended to be more strongly anticorrelated with the subgenual cingulate. They also found that the commonly used targeting method of measuring five centimetres forward from the motor cortex tended to land on sites that were less anticorrelated than sites chosen by some other methods. And they identified a coordinate in the dorsolateral prefrontal cortex where anticorrelation with the subgenual cingulate was strongest across the group, suggesting it as an optimised target. This retrospective finding has since been tested repeatedly. Analyses of data from individual patients in several independent cohorts, including work by Weigand and colleagues in 2018 and by Cash and colleagues in Australia, have found that the strength of anticorrelation between each patient's actual stimulation site and the subgenual cingulate predicts clinical improvement, though the strength of the relationship varies between datasets and some analyses have found it weaker or absent. The effect sizes are modest. In none of these studies does subgenual connectivity explain most of the variation in outcome. But the relationship has been found often enough, in enough independent samples, that it has become the leading hypothesis for why some placements work better than others. Why a few centimetres matter The practical upshot of the network view is that the precise location of the coil within the dorsolateral prefrontal cortex matters. Studies that have mapped connectivity with the subgenual cingulate across the dorsolateral region find that anticorrelation varies considerably over distances of a centimetre or two. A coil placed slightly too far back may stimulate premotor cortex with little connection to mood networks. A coil placed slightly too far forward or too low may stimulate a region with weaker or even positive connectivity to the subgenual cingulate. Individual variation compounds this. The location of the most strongly anticorrelated spot differs from person to person, sometimes by several centimetres. Group-average coordinates, like the one Fox identified, are a reasonable starting point but may miss the individual optimum. And individual brains differ in size, shape and position within the skull, so that the same scalp location overlies different cortical regions in different people. These considerations explain why targeting methods have received so much attention. The earliest clinical method, measuring a fixed distance forward from the motor cortex, ignored individual variation in head size and brain anatomy. Later methods have tried, with increasing sophistication, to account for it. The next chapter examines these methods and the evidence comparing them. Other targets Although the dorsolateral prefrontal cortex dominates, other targets have been explored, and their existence clarifies what the dorsolateral target is doing. The dorsomedial prefrontal cortex, on the upper inner surface of the frontal lobes near the midline, has been targeted in several open and controlled studies, particularly by Jonathan Downar and colleagues in Toronto. It lies deeper than the dorsolateral region, between the hemispheres, and is usually stimulated with a double-cone coil that projects the field downward. Results have been comparable to dorsolateral stimulation in some studies, suggesting that more than one cortical entry point can engage depression-relevant networks. The orbitofrontal cortex, on the underside of the frontal lobes above the eyes, has been targeted with low-frequency stimulation in patients who did not respond to dorsolateral stimulation, with some promising preliminary results. Because it is involved in reward and the evaluation of outcomes, it may be particularly relevant to patients with prominent anhedonia, the loss of capacity for pleasure. The existence of these alternatives supports the network interpretation. If depression reflects disordered activity across a distributed network of cortical and subcortical regions, then any superficial node with strong connections to the relevant deep structures is a candidate target. The left dorsolateral prefrontal cortex is not uniquely privileged. It is the most studied target, the easiest to reach with a standard coil, and the one with the best evidence, but it may not be the best target for every patient. There is also growing interest in matching targets to symptom profiles. Analyses by Siddiqi, Fox and colleagues have suggested that stimulation sites connected to one network may preferentially relieve dysphoric symptoms such as sadness and anhedonia, while sites connected to a different network may preferentially relieve anxious symptoms. If confirmed, this would mean that the best target depends not only on the individual's brain but on the particular pattern of their depression. This is a promising but early line of work, not yet a basis for routine practice. The target as a hypothesis It is worth stepping back to see what kind of claim the choice of target represents. When a clinician places the coil over the left dorsolateral prefrontal cortex, they are acting on a hypothesis: that stimulating this region will, through its connections, alter activity in deep structures involved in mood, and that this alteration, repeated over weeks, will relieve depression. The hypothesis has been refined over thirty years, from the early idea of activating an underactive region to the current idea of modulating a network through its most accessible and best-connected node. Each refinement has made the hypothesis more specific, and therefore more testable, and has made precise targeting more important. This is the first of the booklet's two measurements. If the target is wrong, no amount of stimulation at the correct intensity will produce the intended network effect. If the target is right but the intensity is too low, the stimulation may not engage the network sufficiently. The chapter after next turns to intensity. First, the practical question: given the theoretical target, how does one actually find it on a particular patient's head? Hashtags: #TranscranialMagneticStimulation #TMS #RefractoryDepression #TreatmentResistantDepression #Neuromodulation #RepetitiveTMS #rTMS #DorsolateralPrefrontalCortex #MotorThreshold #TargetingPrecision #StimulationIntensity #ThetaBurstStimulation #iTBS #DeepTMS #ConnectivityGuidedTargeting #SubgenualCingulate #FunctionalConnectivity #SynapticPlasticity #NetworkModulation #IndividualizedDosing #AcuteEfficacy #MaintenanceTMS #RelapsePrevention #ElectroconvulsiveTherapy #FutureOfPrecisionNeuromodulation

Download the Book (PDF): Introduction Stand at a busy urban intersection at five in the evening and count. In a single lane, once the light turns green and the queue begins to move, cars cross the stop line at a remarkably steady rate: one roughly every two seconds, give or take a few tenths. Watch for an hour and you will find that this number barely changes. It does not care whether the drivers are late, whether the road beyond is wide or narrow, or whether the city has just spent a fortune repaving the corridor. It is close to a physical constant of human driving, and it is the single most important number in urban traffic engineering. From that steady two-second rhythm follows almost everything else. If a lane can discharge about 1,800 vehicles in an hour of uninterrupted green, and the lane actually receives green for only 40 percent of the time, then its real capacity is about 720 vehicles an hour, no matter how smoothly the pavement is laid or how many lanes the road has between intersections. The intersection, not the road, decides how much traffic a city street can carry. And because every approach to an intersection must share the same patch of pavement with every other approach, green time given to one direction is green time taken from another. Urban traffic is, at its core, a problem of dividing a fixed resource among competing claims. That is the controlling idea of this book. Urban traffic flow is governed by conflict and time: at every point where movements cross, capacity is fixed by how often and how efficiently the conflict can be resolved, and delay grows explosively as demand approaches that capacity. Signal timing, highway capacity analysis, queueing theory and interchange design are not four separate subjects. They are four views of the same problem. Queueing models explain why the last few percent of capacity cost so much. Capacity analysis measures how close a given intersection is to that cliff. Signal timing and coordination divide the scarce green time so that as little of it as possible is wasted. And modern interchange forms such as the diverging diamond are geometric answers to the same question: if conflicts consume time, remove the conflicts, and the time comes back. Why this framing matters Most people who think about traffic think about speed. A road feels good when it is fast and bad when it is slow. Engineers inherited a version of this instinct, and for decades the default response to congestion was to widen the road. The trouble is that widening the road between intersections rarely touches the actual constraint. A six-lane arterial feeding a signal that gives its through movement only a third of the cycle carries traffic like a two-lane road running continuously. Building the extra lanes without changing the intersection buys queue storage and very little else. The conflict-and-time framing changes which questions get asked. Instead of "how fast can we make this road?" the engineer asks: how many seconds of each cycle are lost to starting and stopping? How many separate phases does this intersection really need, and could geometry eliminate one of them? Are vehicles arriving at the signal while it is green, or are they being stopped and released in clumps that waste green time? How close to capacity is the critical movement, and what happens to delay if demand grows by five percent? Each of those questions has a quantitative answer, and the answers are often surprising. Consider a single example that runs through several chapters. A signalized approach operating at 80 percent of its capacity might impose an average delay of 25 or 30 seconds per vehicle, which most drivers barely notice. Push the same approach to 95 percent and the delay rises perhaps to 40 seconds. Push it to 105 percent and, in a sustained peak, delay does not rise by another ten or fifteen seconds; it climbs without bound, because each cycle leaves a residual queue that the next cycle inherits. The relationship between demand and delay is not a gentle slope but a hockey stick. Almost every practical decision in signal operations, from the choice of cycle length to the justification for an interchange reconstruction, is ultimately an argument about where on that hockey stick a facility sits and how to move it leftward. What the reader will come away with The book is written for intelligent readers who are not traffic engineers: planners, elected officials, students in adjacent disciplines, engineers from other fields, and citizens who have sat at a red light with no cross traffic and wondered what on earth the signal was thinking. It assumes comfort with simple algebra and nothing more. Every formula that appears is explained in words, and every important formula is worked through with numbers so that the reader can see what it does rather than take it on faith. By the end, a reader should be able to do several specific things. They should be able to estimate the capacity of a signalized lane from first principles and explain why a lane's capacity depends on its share of green. They should understand how the Highway Capacity Manual assigns a level of service letter grade to an intersection, what that grade does and does not mean, and why an intersection graded "F" may be operating exactly as intended. They should be able to set out a reasonable fixed-time signal plan for an isolated intersection, including the yellow and all-red intervals and the pedestrian timing that so often governs the result. They should understand how actuated signals respond to detected traffic, how a series of signals is coordinated to produce a green wave, and why green waves work beautifully in some street grids and are almost impossible in others. And they should be able to explain, to a skeptical public meeting if necessary, why a diverging diamond interchange sends drivers briefly onto the left side of the road and why that is safer and more efficient, not less. How the argument unfolds The chapters build on one another. The first establishes the intersection as the network's bottleneck, introducing saturation flow, lost time and the arithmetic of capacity. The second steps back to the physics of traffic streams: the relationship among flow, density and speed, and the shockwaves that carry the effects of a red light upstream through a queue. The third introduces queueing theory, the mathematics that explains why delay rises so sharply near capacity, and derives the delay of a signalized approach from a simple picture of vehicles accumulating during red and discharging during green. With those foundations in place, the fourth chapter turns to measurement: how the Highway Capacity Manual calculates control delay and level of service at signalized and unsignalized intersections, with a full worked example. The fifth chapter shows how to time an isolated signal, from phasing and the critical lane groups to Webster's optimal cycle and the change intervals that keep drivers safe. The sixth explains actuated control, the detector-driven logic that runs most modern signals. The seventh turns to coordination and the green wave, and the eighth to adaptive systems and the performance measures that now allow agencies to watch every signal cycle as it happens. The last chapters take the argument beyond the signalized intersection. The ninth looks at freeway capacity, merges and bottlenecks, where the same queueing logic operates without any signal at all, and where the junction between freeway and arterial at the interchange becomes the hardest problem in the network. The tenth examines interchange forms and their capacity. The eleventh is devoted to the geometric design and operation of the diverging diamond interchange, the clearest modern example of solving a capacity problem by eliminating conflicts rather than by adding pavement. The conclusion draws the threads together and considers what the conflict-and-time view implies for how cities should invest. A note on units and sources. The book uses United States customary units, because the main analytical reference for the field, the Highway Capacity Manual published by the Transportation Research Board, uses them, as does the federal Manual on Uniform Traffic Control Devices. Readers working in metric can convert freely; nothing in the argument depends on the unit system. Where specific thresholds, default values or design guidance are cited, they come from those manuals or from the standard guidance documents of the Federal Highway Administration and the Institute of Transportation Engineers, which are listed in the further reading at the end. Where practice varies among agencies, the text says so rather than pretending to a uniformity that does not exist. Chapter 1. The Intersection Is the Bottleneck A road has two kinds of places. There are links, the stretches of pavement between junctions, where vehicles travel in the same direction and interact only by following and passing one another. And there are nodes, the junctions themselves, where streams of traffic heading in different directions must use the same physical space. On a freeway, links dominate: a vehicle can travel ten miles without crossing anyone's path. On an urban street, nodes dominate: there may be a signalized intersection every quarter mile or less, and at each one the street's traffic must take turns with the traffic of another street. The argument of this chapter is simple but consequential. On an urban network, the capacity of the system is set almost entirely by its intersections, and the capacity of an intersection is set by how much time each movement is allowed to use the shared space and how efficiently it uses that time. Once that is understood, a great deal of urban traffic engineering follows almost mechanically. The queue discharge: saturation flow and headway Begin with a single lane at a signalized intersection. The light is red and a queue of a dozen cars has formed. The light turns green. What happens? The first driver has to notice the change, take a foot off the brake, and accelerate from a standstill. That takes time. The second driver has to wait for the first to begin moving, then begin moving too, with a small additional reaction. By the time the fourth or fifth vehicle crosses the stop line, the queue has stretched out, drivers are moving at a reasonable speed, and each subsequent vehicle crosses the line a fairly constant interval after the one before. That interval is the saturation headway, usually written h. Under good conditions in a through lane, with passenger cars and no interference, it is close to two seconds, and in many North American cities somewhat less, perhaps 1.8 or 1.9 seconds. Field studies dating back to Bruce Greenshields's work in the 1940s show the same pattern repeatedly. The first headway, measured from the start of green to the moment the first vehicle's rear axle crosses the stop line, is the longest, often three and a half to four seconds. The second is shorter, the third shorter still, and by the fourth to sixth vehicle the headways settle around the saturation value. The extra time consumed by the first few vehicles, beyond what they would have needed at the saturation headway, is called start-up lost time. Summed over the first several vehicles it typically amounts to about two seconds. The reciprocal of the saturation headway is the saturation flow rate, written s. If one vehicle crosses the line every two seconds, the lane discharges 3,600 divided by 2, or 1,800 vehicles per hour of green. This is the rate the lane would sustain if the light stayed green forever and the queue never ran out. It is not the lane's capacity, because the light does not stay green forever; it is the maximum rate during the green. The Highway Capacity Manual, the reference document for this kind of analysis in the United States and widely used elsewhere, adopts a base saturation flow rate of 1,900 passenger cars per hour per lane for metropolitan areas of 250,000 people or more, and 1,750 for smaller areas, reflecting the observation that drivers in large cities tend to accept shorter headways. The word "base" matters. Real lanes rarely achieve the base rate, because something about them departs from the ideal: the lane is narrow, there are heavy vehicles in the stream, the approach is on an upgrade, buses stop in the lane, cars park and unpark alongside it, pedestrians conflict with turning vehicles, or the lane is shared between through and turning traffic. The manual applies a series of multiplicative adjustment factors to the base rate to account for each of these conditions. A lane that is ten feet wide instead of twelve, carrying ten percent trucks, on a two percent upgrade, with right-turning traffic yielding to a heavy crosswalk, might end up with an adjusted saturation flow of 1,500 or 1,600 vehicles per hour rather than 1,900. Left turns deserve particular mention because they so often decide an intersection's fate. A left turn made on a protected green arrow, with no opposing traffic, discharges at nearly the rate of a through movement, though the manual applies a modest reduction for the tighter path. A left turn made on a permissive green, yielding to oncoming traffic, behaves completely differently: it can only depart when a gap appears in the opposing stream or when the opposing queue has cleared, and at the end of the phase a few vehicles typically "sneak" through on the yellow. Its effective saturation flow depends on the opposing volume and can be a small fraction of the protected rate. When opposing volumes are high, a permissive left turn may have almost no capacity during the green at all and rely entirely on those end-of-phase sneakers. Lost time and effective green The light does not simply switch from green to red. Between them comes a yellow change interval and often an all-red clearance interval, during which no movement is supposed to enter. Drivers use part of the yellow: some proceed through it because they are too close to stop comfortably. The portion of the yellow and all-red that drivers do not use is called clearance lost time, or sometimes the end lost time. Engineers combine these effects into a single convenient abstraction. Instead of modeling a messy green, yellow and red, with slow start-up at the front and partial use of the yellow at the end, they imagine an idealized "effective green" during which the lane discharges at exactly the saturation rate, and an "effective red" during which it discharges nothing. The effective green is the displayed green plus yellow plus all-red, minus the start-up lost time and minus the clearance lost time. The sum of those two losses is the total lost time per phase, often written t_L. In older editions of the Highway Capacity Manual a default of about four seconds per phase was commonly used; the current method computes it from its components, but four to five seconds per phase remains a reasonable rule of thumb. Lost time is the central villain of signal timing. It is time during which the intersection is controlled by a signal but no one is using it. Every time the signal changes phase, the intersection wastes those seconds. A signal with four phases and a lost time of four seconds per phase loses sixteen seconds every cycle. If the cycle is 80 seconds long, that is twenty percent of all available time, gone. If the cycle is 160 seconds long, it is ten percent. This is the reason that longer cycles carry more traffic: they change phase less often per hour and therefore waste a smaller share of the hour. It is also the reason, as later chapters will show, that longer cycles impose more delay on everyone. The tension between those two facts shapes every signal timing decision. Capacity of a lane group With saturation flow and effective green defined, capacity follows directly. If a lane discharges at s vehicles per hour of green, and receives g seconds of effective green in each cycle of length C seconds, then over an hour it receives g/C of the hour as green, and its capacity is: c = s × (g/C) A through lane with an adjusted saturation flow of 1,800 vehicles per hour that receives 36 seconds of effective green in a 90 second cycle has a capacity of 1,800 × 0.4 = 720 vehicles per hour. The ratio g/C is called the green ratio, and it is the most important lever an engineer has at a signalized intersection. The Highway Capacity Manual groups lanes that are analyzed together, such as two adjacent through lanes, into a lane group, and computes capacity for the group as the sum of its lanes. The ratio of demand flow to capacity, v/c, is called the degree of saturation and usually written X. An approach with 600 vehicles per hour of demand and 720 of capacity has X = 0.83. When X reaches 1.0, the approach is at capacity: on average, exactly as many vehicles arrive in a cycle as can be served. When X exceeds 1.0, vehicles arrive faster than they can be served, and queues grow from cycle to cycle. Note what does not appear in that formula. The width of the road between intersections does not appear. The speed limit does not appear. The quality of the pavement does not appear, except insofar as it affects saturation headway. A lane's capacity at a signalized intersection is its saturation flow times its share of green, nothing more. A city that wants a street to carry more traffic must either give that street a larger share of green, which takes it from the cross street, or raise the saturation flow, which is possible only within narrow limits, or add lanes at the intersection itself. The critical movements and the intersection as a whole An intersection is more than a collection of lanes. Movements that conflict with one another cannot be served at the same time. Northbound through traffic and eastbound through traffic cannot both have green. Northbound left turns and southbound through traffic conflict, which is why permissive left turns must yield. The set of movements that can be served together is called a phase, and the sequence of phases is the phasing plan. Within each phase, some movements need more green than others. Suppose that during the north-south through phase, the northbound lane group carries 700 vehicles per hour on a saturation flow of 3,600 (two lanes at 1,800 each), while the southbound lane group carries 500 on the same saturation flow. The ratio of demand to saturation flow, v/s, is called the flow ratio. It represents the fraction of the hour that the movement would need to be green if there were no lost time. Northbound needs 700/3,600 = 0.194 of the hour; southbound needs 500/3,600 = 0.139. The phase must be long enough to serve the northbound group, which is therefore the critical lane group for that phase. The southbound group will simply have some spare capacity. Add up the critical flow ratios for every phase in the cycle and you have the sum of critical flow ratios, usually written Y. It is the fraction of each hour the intersection would need to be green, collectively, to serve the critical demand if no time were lost to phase changes. If Y is 0.65, then 65 percent of every cycle must be effective green for critical movements. The remaining 35 percent is available for lost time and for spare capacity. From this follows the most useful single equation in signal analysis, the critical degree of saturation: X_c = Y × C / (C − L) where L is the total lost time per cycle. It answers the question: for the intersection as a whole, how close are the critical movements to capacity? With Y = 0.65, lost time of 16 seconds and a 90 second cycle, X_c = 0.65 × 90 / 74 = 0.79. The intersection has some breathing room. If the same intersection had Y = 0.85, X_c would be 1.03 at the 90 second cycle: the critical movements collectively need more time than exists, and no allocation of green among them can avoid queues growing during the peak. Lengthening the cycle to 150 seconds would reduce lost time's share and bring X_c to 0.85 × 150 / 134 = 0.95, just under capacity, at the cost of substantially greater delay for every vehicle and pedestrian. The critical degree of saturation has a valuable property: it is independent of how green is split among phases. It tells you whether any timing plan can serve the demand at the given cycle length. If X_c exceeds 1.0, the problem is not the timing. It is the intersection's geometry, its phasing, or the demand itself. A worked example Consider a four-leg intersection of two arterials, each with two through lanes and an exclusive left-turn lane in each direction, operating with protected left turns. In a typical protected-left scheme the north-south street has a left-turn phase followed by a through phase, and the east-west street does the same, giving four critical phases in sequence. The peak-hour demands, adjusted saturation flows and flow ratios for the movements that turn out to be critical are summarized in Table 1. Table 1. Critical movements in a four-phase worked example. Phase Critical movement Demand (veh/h) Adjusted saturation flow (veh/h) Flow ratio v/s 1 Northbound left 220 1,700 0.129 2 Southbound through 1,050 3,500 0.300 3 Westbound left 160 1,700 0.094 4 Eastbound through 780 3,450 0.226 Sum Y 0.749 Source: illustrative values constructed for this example, using saturation flows in the typical range for adjusted conditions under Highway Capacity Manual methods. The sum of critical flow ratios is 0.749. With four phases at four seconds of lost time each, L = 16 seconds. The minimum cycle length that would just reach capacity, setting X_c to 1.0, is C_min = L / (1 − Y) = 16 / 0.251 = 64 seconds. At that cycle every second of effective green would be fully used during the peak and any random surge would leave a residual queue. A practical design would target X_c of about 0.85 to 0.90, which gives C = L × X_c / (X_c − Y). For X_c = 0.90 that is 16 × 0.90 / 0.151 = 95 seconds, which the engineer would round to 95 or 100. At C = 100, available effective green is 84 seconds. Allocated in proportion to the flow ratios, phase 1 receives 84 × 0.129 / 0.749 = 14.5 seconds, phase 2 receives 33.6, phase 3 receives 10.5 and phase 4 receives 25.3. Each critical movement then operates at the same degree of saturation, about 0.89. The southbound through movement, for instance, has capacity 3,500 × 33.6 / 100 = 1,176 vehicles per hour against demand of 1,050. Three lessons are already visible. First, the left turns, despite carrying only a few hundred vehicles each, consume roughly a quarter of the cycle's effective green, because each requires its own phase with its own lost time. Eliminating one of those phases, by letting a left turn operate permissively or by redesigning the intersection, would release four seconds of lost time per cycle plus whatever green the phase used. Second, the cycle length needed for a comfortable degree of saturation is half again longer than the minimum, a premium paid in delay to buy resilience. Third, the margin is thin. If peak demand on every critical movement grew by ten percent, Y would rise to 0.824 and the 100 second cycle would yield X_c = 0.98. The intersection would be on the edge of failure. Why widening the road between signals does so little The arithmetic of lane-group capacity explains a pattern that frustrates many cities. A corridor is congested, so the road is widened from four lanes to six between intersections. Congestion returns within a few years, sometimes within months. Part of the explanation is induced demand, the well-documented tendency for added road capacity to attract additional traffic; economists Gilles Duranton and Matthew Turner, studying United States metropolitan areas, found vehicle travel rising roughly in proportion to lane miles of major roads. But part of it is simpler. If the widening adds lanes between intersections without adding lanes at the stop lines, or adds a through lane at the stop line but leaves the critical left turn constrained to a single lane with a protected phase, the critical movements barely change. The lane-group capacities of the critical movements set the intersection's limit; added pavement elsewhere only stores longer queues. Conversely, modest interventions at the intersection itself often deliver more than major widening. Adding a second left-turn lane to a heavily used protected left can halve that movement's flow ratio, freeing green for everyone. Converting a four-phase intersection to three phases by allowing one pair of lefts to run permissively when opposing volumes permit saves a full phase of lost time. Moving a bus stop from the near side to the far side of an intersection removes a source of lane blockage that depresses saturation flow. Prohibiting a low-volume left turn and redirecting it around the block can remove a phase entirely. These are the levers engineers reach for first, because the arithmetic says they matter most. The rest of this book is built on the vocabulary of this chapter: saturation flow, lost time, effective green, lane-group capacity, degree of saturation, critical flow ratio. With them in hand, the next step is to understand what traffic does when it is not at a stop line, as a moving stream whose density and speed propagate effects up and down the road, and then to see why the closeness of X to 1.0 matters so much more than intuition suggests. Chapter 2. Traffic Streams, Shockwaves and the Physics of the Queue A signal does not act only at its stop line. When it turns red, the effect travels backward through the arriving traffic, stopping vehicles hundreds of feet away. When it turns green, a wave of motion propagates back through the queue, and a driver near the rear may wait twenty seconds after the light has changed before their car begins to roll. To understand signals, and later freeway bottlenecks and interchange spacing, it helps to understand traffic as a stream with physical properties of its own. This chapter sets out those properties and the simple but powerful theory that connects them. Flow, density and speed Three quantities describe a traffic stream at any point. Flow, q, is the number of vehicles passing a point per unit time, usually expressed in vehicles per hour. Density, k, is the number of vehicles occupying a length of road at an instant, usually expressed in vehicles per mile per lane. Speed, v, is the average speed of the vehicles, more precisely the space-mean speed, the average taken over the vehicles occupying a length of road rather than over those passing a point. These three are linked by an identity that holds for any stream in steady conditions: q = k × v If there are 40 vehicles per mile in a lane and they travel at 45 miles per hour, then 1,800 vehicles pass a point each hour. The identity is just bookkeeping, but it has a striking consequence when combined with the observation that speed depends on density. Drivers travel fast when the road ahead is empty and slow down as it fills. Bruce Greenshields, in a study published in 1935, proposed that speed falls linearly with density, from a free-flow speed when the road is empty to zero when vehicles are packed bumper to bumper at the jam density. Real data are not perfectly linear, and modern models differ, but the qualitative shape of Greenshields's relationship has held up for ninety years. Substituting a falling speed into q = k × v gives a flow that rises from zero at zero density, reaches a maximum at some intermediate density, and falls back to zero at jam density. That maximum is the capacity of the road, and the density at which it occurs is the critical density. Two features of this flow-density relationship, usually called the fundamental diagram, drive almost everything else. First, for any flow below capacity there are two possible states: an uncongested state with low density and high speed, and a congested state with high density and low speed. A freeway lane carrying 1,500 vehicles per hour could be doing so at 60 miles per hour with 25 vehicles per mile or at 15 miles per hour with 100 vehicles per mile. The flow is the same; the experience is completely different. Second, in the congested branch, adding vehicles reduces flow. Packing more cars into a jammed road does not move more of them. Modern traffic science generally represents the fundamental diagram as roughly triangular rather than parabolic. On the uncongested side, speed stays near the free-flow speed until close to capacity, so flow rises almost linearly with density. On the congested side, flow falls almost linearly from capacity to zero at jam density. This triangular form, associated with Carlos Daganzo and with Gordon Newell's simplified car-following theory, is not merely a convenience. It reflects how drivers actually behave: in congestion, each driver keeps roughly a fixed spacing plus a reaction time behind the vehicle ahead, and that rule produces a straight congested branch. Typical values make the picture concrete. A freeway lane might have a free-flow speed of 65 to 70 miles per hour, a capacity of 2,200 to 2,400 passenger cars per hour, a critical density of 35 to 45 passenger cars per mile, and a jam density of around 190 to 210 vehicles per mile, corresponding to a spacing of about 25 to 28 feet between front bumpers in a stopped queue. At a signalized intersection, the jam density of a stopped queue is similar, and the saturation flow plays the role of capacity for the discharging stream. Shockwaves: how a red light reaches upstream When two traffic states meet, the boundary between them moves. That moving boundary is called a shockwave, and its speed follows from conservation of vehicles: vehicles cannot appear or vanish at the boundary, so the number entering it from one side must equal the number leaving it on the other. The result, derived independently by Michael Lighthill and Gerald Whitham in 1955 and by Paul Richards in 1956 in what is now called the LWR model, is that the speed of the boundary between state A and state B is: w = (q_B − q_A) / (k_B − k_A) the change in flow divided by the change in density across the boundary. On a flow-density graph, it is the slope of the line joining the two states. Apply this to a signal turning red. Upstream, traffic arrives in an uncongested state: say 900 vehicles per hour in a lane at a density of 25 vehicles per mile, which corresponds to a speed of 36 miles per hour. At the stop line, vehicles stop and form a queue at jam density, say 200 vehicles per mile, with zero flow. The boundary between arriving traffic and the stopped queue is the back of the queue. Its speed is (0 − 900) / (200 − 25) = −5.1 miles per hour. The negative sign means the boundary moves upstream, against the direction of traffic. The back of the queue grows backward at about 5 miles per hour, or roughly 7.5 feet per second. After 40 seconds of red it has reached 300 feet from the stop line, about twelve vehicles back. Now the signal turns green. The queue at the front begins to discharge at saturation: say 1,800 vehicles per hour at a density, in the moving discharge state, of about 50 vehicles per mile. The boundary between the stopped queue and the discharging traffic is the start-up wave, the point at which vehicles begin moving. Its speed is (1,800 − 0) / (50 − 200) = −12 miles per hour, again moving upstream, but more than twice as fast as the queue was growing. Roughly 17.6 feet per second. A driver 300 feet back will begin to move about 17 seconds after the start of green. Meanwhile the back of the queue keeps growing upstream at 5 miles per hour as new vehicles arrive. The start-up wave is chasing it at 12 miles per hour. They meet when the start-up wave has caught the back of the queue, at which point the queue has fully begun moving. Behind that point, arriving vehicles can proceed without stopping, though the discharging platoon ahead of them continues to clear at saturation until it has passed the stop line. This simple graphical exercise, carried out on a time-space diagram showing distance from the stop line against time, gives engineers a powerful tool. It predicts the maximum queue length, which determines whether a queue will spill back and block an upstream intersection or a driveway. It predicts how long it takes a queue to clear, which tells you how long the green must be. And it explains why queues at closely spaced intersections interact. If the back of queue at one signal reaches the upstream intersection before the start-up wave arrives, vehicles in the upstream intersection cannot leave, and the upstream intersection's capacity collapses even though its own signal is green. This condition, often called spillback or blocking, is one of the main ways urban networks fail, and it is exactly the condition that the timing and coordination methods of later chapters aim to prevent. The time-space diagram Almost everything in signal coordination and interchange operation is reasoned out on the time-space diagram, and it is worth describing it carefully because it will reappear throughout the book. Imagine a horizontal axis of time, in seconds, and a vertical axis of distance along a road. A vehicle traveling at constant speed is a straight line rising to the right; the steeper the line, the faster the vehicle. A stopped vehicle is a horizontal line. Each signal along the road is a horizontal band at its location, colored green or red as time passes. A vehicle's trajectory, rising across the diagram, must not cross a signal band while it is red; if it reaches a signal during red, the trajectory goes flat until the green begins, then rises again after a short acceleration. On this diagram, shockwaves are also straight lines, but they slope the other way: the back of a growing queue at a signal is a line descending from the stop line backward in distance as time increases, and the start-up wave is a steeper descending line from the moment the light turns green. The triangle formed between the red interval, the queue growth line and the start-up line is the region occupied by stopped vehicles. Its area, in vehicle-seconds per foot, relates directly to the total delay incurred by the queue. Engineers designing coordinated signal systems draw the time-space diagram for the whole corridor and look for a "band" of parallel trajectories that can pass through every signal on green. The width of that band in seconds is the bandwidth, and maximizing it was the focus of signal coordination for half a century. Chapter 7 develops that idea in detail. Capacity drop and the fragility of maximum flow One feature of real traffic does not appear in the simple fundamental diagram but turns out to be important: the maximum flow a road carries before it breaks down is often higher than the flow it discharges once a queue has formed. On freeways this is called the capacity drop, and measured values typically fall in the range of a few percent to around fifteen or twenty percent. A freeway bottleneck might carry 2,300 vehicles per hour per lane for a few minutes just before breakdown, then settle to discharging 2,050 per lane from the queue once congestion forms. The mechanism is debated but plausibly rooted in driver behavior. In free flow near capacity, drivers maintain short headways at high speed. Once they have been forced to slow down and then accelerate out of a queue, they tend to accept longer headways and accelerate more gently, so the discharge rate is lower. Lane changing in the vicinity of the bottleneck also disrupts flow. The capacity drop matters for two reasons. It means that preventing breakdown is worth more than the arithmetic of capacity would suggest: once a queue forms, the road loses a portion of its throughput for as long as the queue lasts. And it provides the theoretical justification for ramp metering and for some forms of signal control that hold traffic back in order to keep a downstream bottleneck just below breakdown. At signalized intersections an analogous phenomenon appears when queues from downstream spill back into the intersection. The saturation flow assumed in the capacity calculation simply does not materialize when the receiving road is full. Platoons and their dispersion Traffic leaving a signal does not arrive at the next signal as a uniform stream. It arrives as a platoon, a dense group of vehicles released during the green, followed by a sparse trickle of vehicles that entered the road from side streets and driveways during the red. How tightly that platoon holds together as it travels determines how well downstream signals can be coordinated with it. Platoons disperse because drivers choose different speeds. Faster drivers pull ahead, slower ones fall behind, and over a distance of half a mile or more the platoon spreads out. The classic model of this process, developed by Denis Robertson at the United Kingdom's Transport and Road Research Laboratory and embedded in the TRANSYT signal optimization program he created in the late 1960s, represents the downstream flow as a smoothed and delayed version of the upstream flow, using a dispersion factor calibrated to the road's conditions. On a short block with few driveways, the platoon arrives almost intact. On a long suburban arterial with many access points, it may arrive as an almost uniform stream. This is a practical matter. Coordination produces large benefits when signals are close together, generally less than half a mile apart, and platoons remain tight. On long, widely spaced corridors with heavy mid-block access, platoons disperse and coordination's benefits shrink, which is why some agencies run signals on such corridors in free, uncoordinated actuated mode for parts of the day. Chapter 7 returns to the question of when coordination pays. What the stream view adds The intersection analysis of Chapter 1 treats each approach as a single point with a saturation flow and a green time. The stream view adds space. It tells us how long queues become, how fast they grow and clear, whether they will reach the next intersection upstream, and how arrivals at one signal are shaped by departures from the last. Without it, an engineer could compute capacity correctly and still design a corridor that fails because queues block the intersections that feed them. But the stream view, in its basic deterministic form, misses something essential. It treats arrivals as smooth and predictable. Real arrivals are random. Two cycles with the same average demand do not deliver the same number of vehicles, and that randomness turns out to matter enormously as demand approaches capacity. The mathematics of randomness at a service point is queueing theory, and it is the subject of the next chapter. Chapter 3. Why Delay Explodes Near Capacity: Queueing Models Every driver knows the experience. On most days, a particular intersection is a minor nuisance: a wait of half a minute, perhaps less. Then one day traffic is only slightly heavier, maybe because of a school event or an incident on a parallel road, and the same intersection becomes a disaster, with queues stretching back several blocks and waits of five minutes or more. The demand did not double. It may have risen by ten percent. Yet the delay multiplied several times over. This is not a quirk of a badly timed signal. It is a mathematical property of any system in which customers arrive at a server with limited capacity, and it is the reason traffic engineers obsess about the degree of saturation. This chapter explains that property, starting with the general theory of queues and then applying it to the signalized intersection, where it yields the delay formulas at the heart of the Highway Capacity Manual. The anatomy of a queue A queueing system has three essential elements. Customers arrive according to some pattern, the arrival process. They are served by one or more servers, each with a service time that may be fixed or variable, the service process. And they wait in line according to some rule, the queue discipline, which for traffic is almost always first come, first served within a lane. Queueing theorists use a compact notation introduced by David Kendall in 1953 to describe these systems. An M/M/1 queue has Markovian, meaning random and memoryless, arrivals following a Poisson process; Markovian, meaning exponentially distributed, service times; and one server. An M/D/1 queue has random arrivals but deterministic, fixed service times. A D/D/1 queue has both arrivals and service perfectly regular. The key parameter of any single-server queue is its utilization, usually written ρ: the arrival rate divided by the service rate. If cars arrive at a toll booth at 400 per hour and the booth can process 500 per hour, ρ = 0.8. The booth is busy 80 percent of the time on average. Utilization in queueing theory plays exactly the role that the degree of saturation X plays in signal analysis. The D/D/1 queue: no randomness Start with the simplest case, where vehicles arrive at perfectly regular intervals and are served at perfectly regular intervals. If ρ is less than 1, every vehicle arrives to find the server idle, because the previous one has already finished. There is no queue and no delay. If ρ exceeds 1, vehicles arrive faster than they can be served and the queue grows linearly without limit. The relationship between utilization and delay is a step: zero below capacity, unbounded above it. This model is unrealistic for a continuous server, but it captures something true about signalized intersections, which have an additional complication: the server is periodically switched off. Even with perfectly regular arrivals, vehicles arriving during red must wait. This produces a deterministic delay that exists at any level of demand, which Chapter 2 visualized as the triangle on the time-space diagram, and which the Highway Capacity Manual calls uniform delay. The M/M/1 queue: randomness and the hockey stick Now let arrivals be random, as they are on any real road where vehicles come from many origins on their own schedules. Arrivals following a Poisson process have a constant average rate but irregular spacing: sometimes three come in a cluster, sometimes there is a long gap. When several arrive close together, a queue forms even if the server is, on average, much faster than the arrival rate. During a long gap, the server sits idle, and that idle time is lost forever; it cannot be saved up to serve the next cluster. The classic result for an M/M/1 queue is that the average number of customers in the system is ρ/(1 − ρ), and the average time each spends in the system is 1/(μ − λ), where λ is the arrival rate and μ the service rate. The average time waiting in the queue before service begins is ρ/(μ − λ). The term (1 − ρ) in the denominator is the source of the hockey stick. Look at the average number in the system at different utilizations. At ρ = 0.5 it is 1. At 0.8 it is 4. At 0.9 it is 9. At 0.95 it is 19. At 0.99 it is 99. Each step toward capacity adds more queue than the one before. Moving from 80 to 90 percent utilization more than doubles the queue; moving from 90 to 95 percent doubles it again. And at 100 percent the formula breaks down entirely, because the steady-state queue is infinite. This is not specific to traffic. Exactly the same mathematics explains why hospital emergency departments and call centers degrade sharply as they approach full utilization, and why computer networks are engineered to keep links well below their nominal bandwidth. Any system with random arrivals and a fixed capacity behaves this way. The practical lesson is universal: the last few percent of capacity are extraordinarily expensive in waiting time, and a system designed to run at 98 percent utilization will spend much of its life in severe congestion. The M/D/1 queue and the role of service variability At a signal or a toll booth, service is not exponentially distributed: once the queue is moving, vehicles depart at a fairly steady saturation headway. The M/D/1 model, with random arrivals and fixed service times, is closer to reality. Its average wait in queue is ρ / [2μ(1 − ρ)], exactly half that of the M/M/1 queue at the same utilization. The general result, the Pollaczek–Khinchine formula, shows that average waiting time is proportional to (1 + Cs²)/2, where Cs is the coefficient of variation of service time. Fixed service halves the queue relative to exponential service; but the (1 − ρ) term remains in the denominator, and the hockey stick remains. The lesson for traffic is twofold. Reducing variability helps: anything that makes arrivals more regular, such as metering traffic or coordinating upstream signals so that vehicles arrive in predictable platoons, or that makes service more uniform, such as removing lane blockages by parking maneuvers and stopping buses, lowers delay at any given utilization. But no amount of regularity fully escapes the explosive growth of delay near capacity in a system with any randomness at all. From steady state to the peak hour Classic queueing theory deals with steady state: what happens after a system has run at constant rates for a very long time. Traffic peaks are not like that. Demand rises over an hour or so, stays high for a period, and falls again. During that period the queue does not reach the steady state the formulas describe; it builds gradually, and if demand falls before the queue grows very large, the actual delay may be far less than the steady-state formula predicts. Conversely, when demand exceeds capacity for a period, steady-state theory says nothing useful at all, because there is no steady state; but the queue grows at a finite rate, and the delay over a finite period is finite. Traffic engineers dealt with this gap in two stages. In 1958 F. V. Webster, at the United Kingdom's Road Research Laboratory, combined deterministic signal delay with a steady-state random component and a correction term fitted from simulation, producing a formula for average delay at a fixed-time signal that is still taught. It works well for degrees of saturation up to about 0.9 but, like all steady-state formulas, goes to infinity at X = 1. The second stage was the coordinate transformation method, developed in the 1970s and 1980s by researchers including Rahmi Akçelik in Australia and Kimber and Hollis in the United Kingdom. Its idea is elegant. Below capacity, use a steady-state queueing formula. Well above capacity, use the deterministic overflow model, in which the queue grows linearly at the rate (v − c) for the duration of the analysis period. Then construct a smooth curve that approaches the steady-state formula at low X and the deterministic line at high X. The result gives finite, plausible delays at and above capacity for an analysis period of given length. The incremental delay term in the Highway Capacity Manual is a formula of exactly this kind. Delay at a signalized approach The Highway Capacity Manual's control delay for a signalized lane group is built from three terms. Control delay is the extra time a vehicle spends because of the signal, compared with traveling at its desired speed without the signal: the time spent decelerating, stopped, moving up in the queue and accelerating back to speed. The first term, uniform delay, d1, is the deterministic delay of a vehicle arriving at a uniform rate on a signal with fixed red and green. From the triangle of queue growth and discharge, it can be shown to equal: d1 = 0.5 × C × (1 − g/C)² / [1 − min(1, X) × g/C] where C is the cycle length, g the effective green, and X the degree of saturation. This is the basic price of having a signal at all. Notice its structure. It is proportional to the cycle length, so doubling the cycle roughly doubles uniform delay. It rises steeply as the green ratio falls, because the red interval in which vehicles accumulate grows. And it depends only modestly on X, because in a deterministic world below capacity every queue clears in every cycle. The second term, incremental delay, d2, accounts for random arrivals and for oversaturation. In the form adopted by the manual: d2 = 900 × T × [(X − 1) + √((X − 1)² + 8kIX / (cT))] where T is the analysis period in hours (usually 0.25, a fifteen-minute peak), c is the lane group capacity in vehicles per hour, k is a factor reflecting the controller type (0.5 for pretimed control, smaller for well-tuned actuated control), and I is an upstream filtering factor reflecting how much upstream signals have regularized arrivals (1.0 for an isolated intersection with random arrivals). When X is well below 1, the first term inside the brackets is negative and nearly cancels the square root, so d2 is small. When X is well above 1, the expression approaches 1,800 × T × (X − 1) seconds, which is exactly the average delay of the deterministic overflow queue over the analysis period. The third term, d3, is initial queue delay, which applies when a queue left over from a previous period is still present at the start of the analysis period. It is zero for an approach that begins the period without a residual queue. Earlier editions of the manual also multiplied d1 by a progression adjustment factor reflecting how favorably coordinated arrivals align with the green; current editions instead compute uniform delay from an arrival-pattern-based calculation that accounts for the proportion of vehicles arriving on green. The principle is the same either way: good coordination reduces delay by sending vehicles to the signal while it is green. The hockey stick in numbers To see what these formulas mean, take the southbound through lane group from the worked example in Chapter 1: a 100-second cycle, 33.6 seconds of effective green, a capacity of 1,176 vehicles per hour, a fifteen-minute analysis period, pretimed control and random arrivals. Hold the timing fixed and vary the demand. The resulting delays appear in Table 2. Table 2. Control delay against degree of saturation for a fixed signal timing. Degree of saturation X Demand (veh/h) Uniform delay d1 (s) Incremental delay d2 (s) Control delay (s/veh) 0.60 706 27.6 2.3 29.9 0.80 941 30.1 5.8 35.9 0.89 1,047 31.4 10.3 41.7 0.95 1,117 32.4 16.7 49.1 1.00 1,176 33.2 26.2 59.4 1.10 1,294 33.2 58.1 91.3 1.20 1,411 33.2 98.4 131.6 Source: calculated by the author with the uniform and incremental delay equations of the Highway Capacity Manual, using C = 100 s, g = 33.6 s, c = 1,176 veh/h, T = 0.25 h, k = 0.5 and I = 1.0, with no initial queue. The pattern is the hockey stick again, now with real units. Going from 60 to 80 percent of capacity adds six seconds of delay. Going from 80 to 95 percent adds thirteen. Going from 95 to 110 percent, a further fifteen percentage points, adds forty-two. And this is for only a fifteen-minute peak. If the oversaturation persisted for an hour, the incremental delay at X = 1.2 would be four times larger, because the queue would keep growing for four times as long. Two features of the table are worth dwelling on. First, even at low demand, delay is substantial: almost thirty seconds per vehicle at 60 percent of capacity. That is the uniform delay, the price of the red light, and it is controlled mostly by the cycle length and the green ratio, not by demand. An engineer who wants to reduce delay at a lightly used intersection should shorten the cycle, not add lanes. Second, at high demand, the random and overflow components dominate, and they are controlled by the degree of saturation. An engineer who wants to reduce delay at a heavily used intersection must lower X, by adding capacity to the critical movements, lengthening their green, or reducing demand. Queue length and storage Delay is what drivers feel; queue length is what determines whether the network fails. A queue that is too long for its storage space blocks something: an upstream intersection, a left-turn bay entrance, a driveway, or on an off-ramp the freeway mainline itself. The Highway Capacity Manual provides methods to estimate the back-of-queue length, and engineers typically design storage for a high percentile, often the 95th percentile queue, rather than the average, because the average queue is exceeded in roughly half of all cycles. The same queueing mathematics applies. At moderate degrees of saturation, the 95th percentile queue may be roughly twice the average, because random fluctuation produces cycles with substantially more arrivals than usual. Near capacity, cycles in which the queue fails to clear become common, and the residual queue carries into the next cycle, so queue lengths grow and become highly variable. A left-turn bay sized for the average queue will overflow in a large fraction of peak cycles, and when it overflows, the left-turners spill into the adjacent through lane and block it, reducing the through movement's capacity and spreading the failure. This is one of the most common real-world failure mechanisms, and one of the most avoidable. A left-turn bay that is fifty feet too short can cost an intersection more capacity than any timing adjustment can recover. What queueing theory tells the engineer The practical consequences of this chapter can be stated plainly. Design for a degree of saturation comfortably below 1.0, typically 0.85 to 0.95 for critical movements in the peak, because the delay curve is steep above that range and the system has no reserve to absorb fluctuations. Reduce variability where possible, because regular arrivals queue less than random ones at the same volume. Keep cycles as short as capacity permits, because uniform delay scales with the cycle. Size storage for high-percentile queues, not averages. And recognize that a small reduction in demand at a near-saturated intersection, whether through a parallel route, a transit improvement or a modest change in land use, can produce a disproportionately large reduction in delay, for exactly the same reason that a small increase produces a disproportionately large rise. That last point cuts both ways in public debate. It explains why a closed lane or a new development can cause severe congestion out of proportion to the traffic it adds. It also explains why modest demand management, which may look trivial in percentage terms, can be among the most cost-effective congestion relief available. The formulas of this chapter are the engine of the Highway Capacity Manual's method for grading intersections. The next chapter shows how that method works, how delay is translated into a level of service letter, and why the letter grade, used carelessly, can mislead. Hashtags: #TrafficEngineering #SignalTiming #HighwayCapacity #QueueingModels #SaturationFlow #SaturationHeadway #EffectiveGreen #LostTime #LaneGroupCapacity #DegreeOfSaturation #CriticalFlowRatio #SignalPhasing #WebsterCycle #ActuatedSignalControl #SignalCoordination #GreenWave #HighwayCapacityManual #LevelOfService #TrafficFlowTheory #FundamentalDiagram #ShockwaveAnalysis #QueueSpillback #ControlDelay #DivergingDiamondInterchange #FutureOfTrafficEngineering

Download the Book (PDF): Introduction A woman in her late thirties arrives at a community mental health clinic with a referral letter that lists three diagnoses: generalized anxiety disorder, recurrent major depression and insomnia. She has tried four antidepressants. She describes herself as "jumpy," says she cannot sit with her back to a door, and mentions, almost in passing, that she sometimes "goes away" during arguments with her partner and comes back with no memory of what was said. Nobody has asked her what happened to her. When the question is finally put, plainly and without pressure, the answer takes ten minutes to arrive and reorganizes everything in the file: a violent assault at nineteen, years of nightmares she assumed everyone had, a body that braces for impact whenever a car door slams in the street. This booklet is written for the gap that her file reveals. Most psychiatric services were built around a model in which symptoms cluster into disorders, disorders respond to particular drugs or therapies, and the history of what a person has lived through is background information rather than the organizing fact. For a large share of patients that model works poorly, because their symptoms are not a random cluster. They are the predictable output of a defensive system that did exactly what it evolved to do under conditions of overwhelming threat and then failed to stand down. The controlling idea The argument of this booklet is straightforward to state. Post-traumatic stress disorder and its dissociative relatives are best understood as disorders of threat learning and threat regulation, with identifiable signatures in the brain and body, and clinical care improves when it is organized around that understanding. Three practical consequences follow. First, the assessment of almost any psychiatric patient should include a careful, non-intrusive inquiry into trauma and into the bodily and dissociative signs that often accompany it. Second, the treatments with the strongest evidence, including Eye Movement Desensitization and Reprocessing (EMDR), work by changing how trauma memories are stored and retrieved, not by teaching people to tolerate an unchanging injury, and they can be delivered safely to far more patients than are currently offered them. Third, the body is not an afterthought. Arousal, breathing, sleep, posture, pain and the capacity to sense internal states are where trauma lives for many patients, and therapies that work through the body deserve a place in care, assessed with the same evidential standards as everything else. Each of these claims is modest when stated alone. Taken together, they amount to a substantial restructuring of how psychiatric care is delivered: a different intake interview, different expectations about what a diagnosis means, different decisions about who gets referred for trauma-focused therapy and when, and a different attitude toward the patient whose distress shows up as a racing heart or a vacant stare rather than as a narrative. What the neuroscience can and cannot do There is a risk in writing about the neurobiology of trauma, and it is worth naming at the start. Over the past three decades a popular account has formed, built from genuine research but simplified in the retelling: the amygdala is a smoke alarm stuck in the on position; the prefrontal cortex goes offline; trauma is stored in the body rather than the mind; the "thinking brain" cannot reach the "survival brain," so talking cannot help. Parts of this account are well supported. Parts are metaphors that have hardened into claims. Parts are simply wrong. The research does show that people with PTSD, as a group, tend to show heightened responses in the amygdala and related threat-detection regions when exposed to trauma reminders and to fearful faces, and reduced engagement of medial prefrontal regions that normally help suppress learned fear. It shows that the ability to recall that a threat is no longer present, which neuroscientists call extinction recall, is impaired. It shows altered stress hormone dynamics, disrupted sleep architecture, elevated inflammatory markers and reduced variability in heart rhythm. It shows that a subgroup of patients respond to reminders not with alarm but with shutdown, emotional numbing and a sense of unreality, and that this subgroup has a different pattern of brain activity. The research does not show that every patient fits the group average. Brain imaging findings in PTSD are statistical differences between groups with considerable overlap; no scan can diagnose the disorder in an individual. It does not show that trauma memories are stored somewhere other than the brain, or that verbal therapies cannot reach them. On the contrary, the best-supported treatments are largely verbal and work extremely well for many people. And it does not show that any particular somatic technique, however intuitively appealing, is effective simply because it fits a neurobiological story. The value of the neuroscience for a clinician is therefore not that it provides diagnostic tests or replaces clinical judgment. Its value is that it explains why the symptoms take the shape they do, why some treatments work and others do not, why a patient who is dissociating cannot learn in a session, and why the conventional instruction to "calm down and think about it rationally" is close to useless for someone whose defensive physiology is fully engaged. It gives clinicians a working model of the patient in front of them, and a working model is what turns a list of techniques into coherent care. Who this booklet is for The intended reader is an intelligent professional or student who works with, or is preparing to work with, people affected by severe stress: psychiatrists and psychiatric trainees, psychologists, nurses, social workers, general practitioners, counselors, and administrators who make decisions about how services are organized. It assumes no prior training in neuroscience. Terms are defined as they arise, and a short glossary at the end collects the ones that recur. The booklet is not a treatment manual. No one should begin practicing EMDR or any other trauma-focused therapy on the strength of a book; each requires supervised training. What the booklet aims to provide is the understanding that should precede and accompany that training: a clear picture of what trauma does to the brain and body, what the evidence says about treatments, and how a service might be reorganized to put that knowledge to use. The route through the subject The chapters build in a deliberate order. The first establishes how the healthy stress response works, because PTSD is not the presence of an alien process but the persistence of a normal one. The second and third examine the two halves of the most influential neural model of PTSD: an amygdala-centered threat system that fires too readily, and a prefrontal regulatory system that fails to restrain it. The fourth turns to memory, explaining why trauma memories are fragmented, intrusive and stubbornly present-tense, and why the discovery that memories can be modified on retrieval matters so much for treatment. The fifth chapter is devoted to dissociation, the most neglected and most misunderstood aspect of traumatic stress, and to the dissociative subtype of PTSD. The sixth widens the view from the brain to the whole body: the stress hormones, autonomic nervous system, immune system, sleep and the long-term health consequences of severe and early adversity. The second half of the booklet is clinical. The seventh chapter reconsiders assessment and diagnosis, including the distinction between PTSD and complex PTSD introduced in the World Health Organization's ICD-11. The eighth surveys the full treatment landscape, from cognitive and exposure-based therapies to medications and the experimental compounds that have drawn so much attention. The ninth and tenth are devoted to EMDR: first what it is, how it developed and what is known about why it works, and then the practical business of integrating it into clinical services, including how to handle dissociation, complex histories and the long-running debate about whether patients need extensive stabilization before trauma processing. The eleventh assesses body-based therapies with an even hand, separating what has been tested from what has merely been asserted. The twelfth asks what a psychiatric service looks like when it is built from the ground up to recognize the physiological impact of severe stress. The patient described at the start of this introduction did well, as it happens. Once the trauma was named, her symptoms made sense to her for the first time, which was itself a relief. She completed a course of trauma-focused therapy, slept through the night within three months and stopped "going away" during arguments. Her case is a composite, as are all the clinical illustrations in this booklet, but the pattern it describes is common enough that every clinician will recognize it. The aim of what follows is to make that pattern easier to see, and easier to act on, in the first appointment rather than the fifth year. Chapter 1. The Defensive System: How the Body Meets Threat Every account of what goes wrong in post-traumatic stress has to begin with what goes right in ordinary danger. The brain and body possess a defensive system refined over hundreds of millions of years of vertebrate evolution, and nearly every symptom of PTSD is a component of that system operating outside its proper time and place. Hypervigilance is threat detection that will not switch off. Startle is an orienting reflex set to a hair trigger. Nightmares are a memory system rehearsing danger. Emotional numbing and dissociation are versions of the immobility responses that prey animals display when escape has failed. To understand the disorder, a clinician first needs to understand the machinery. The architecture of a threat response When a person encounters danger, two broad families of response unfold in parallel. The first is fast and neural. Sensory information about the threat reaches the amygdala, a pair of almond-shaped clusters of nuclei deep in each temporal lobe, through more than one route. Some of that information arrives quickly and in coarse form; some arrives after more elaborate processing in sensory and association cortex. The amygdala's central nucleus projects to the hypothalamus and brainstem, setting off the sympathetic nervous system within a fraction of a second. Heart rate and blood pressure rise, breathing quickens, the pupils dilate, blood is shunted toward the large muscles, digestion slows, and the skin's sweat glands activate. A small cluster of neurons in the brainstem called the locus coeruleus, the brain's principal source of noradrenaline (norepinephrine), floods the forebrain with that neurotransmitter, sharpening attention toward the threat and away from everything else. The second family is slower and hormonal. The hypothalamus releases corticotropin-releasing hormone, which signals the pituitary gland to release adrenocorticotropic hormone into the bloodstream, which in turn prompts the adrenal cortex to produce cortisol. This chain, the hypothalamic-pituitary-adrenal or HPA axis, takes minutes rather than milliseconds to reach full effect. Cortisol mobilizes glucose, modulates the immune system and, crucially, acts back on the brain. It helps consolidate memories of emotionally significant events, and through receptors in the hippocampus, prefrontal cortex and hypothalamus it eventually shuts the stress response down. The HPA axis is therefore not only an accelerator but also a brake, and its negative feedback is one of the mechanisms by which a healthy organism returns to baseline once danger has passed. The neuroscientist Joseph LeDoux, whose work on fear conditioning in rats did much to establish the amygdala's role, has in recent years argued that we should be careful about calling this machinery the "fear system." His point is that these circuits detect threat and organize defensive behavior and physiology, and they can do so without any conscious experience of fear at all. The subjective feeling of being afraid is built separately, from the brain's interpretation of what the body and the defensive circuits are doing, together with memory and context. This distinction matters clinically. A patient with PTSD may show a full defensive response to a reminder, with racing heart, sweating and muscle tension, while reporting that she feels nothing in particular, or feels irritable, or feels ashamed. The defensive physiology and the conscious emotion are related, but they are not the same thing, and treatments may affect one without the other. The imminence continuum The popular phrase "fight or flight," coined by the physiologist Walter Cannon in the early twentieth century, captures only part of the defensive repertoire. Research on how animals respond to predators, associated particularly with the work of Michael Fanselow and colleagues, describes a sequence of responses that shift as a threat moves closer. The framework is often called the predatory imminence continuum, and it maps remarkably well onto what trauma survivors report. When danger is possible but not yet present, an animal shows heightened vigilance and cautious exploration: it scans the environment, approaches slowly, avoids open ground. In humans this resembles the state of anxious apprehension, a background sense that something might go wrong. When a threat is detected but still at a distance, the characteristic response is freezing, an attentive immobility in which the body is primed for action but motionless. Heart rate often slows briefly in this state, a phenomenon called fear bradycardia, while muscles remain tense and senses are acutely focused. Freezing reduces the chance of being detected and buys time to assess the situation. When the threat closes in, the animal breaks into active defense: flight if escape is possible, fighting if it is cornered. And when the predator makes contact and escape has failed, many species display a final response called tonic immobility, a profound, involuntary paralysis with reduced responsiveness, sometimes accompanied by analgesia and a sense of detachment. Some researchers describe a further response, collapsed immobility or "flag," in which muscle tone is lost, heart rate and blood pressure fall, and the person may feel faint or actually lose consciousness. The evidence for collapse as a distinct defensive state in humans is less systematic than for freezing and tonic immobility, but survivors describe it consistently enough that clinicians should recognize it. Table 1 sets these responses side by side. Table 1. Defensive responses across the threat imminence continuum. Stage of threat Typical response Body state Common human experience Threat possible, not detected Vigilance, risk assessment Mild arousal, scanning Unease, apprehension, hypervigilance Threat detected at distance Attentive freezing Muscles tense, heart briefly slows, senses sharpened "Stopped dead," intense focus, waiting Threat close, escape possible Flight or fight High sympathetic arousal, rapid heart, rapid breathing Panic, rage, urge to run or strike Contact, escape impossible Tonic immobility Involuntary paralysis, reduced pain, tremor possible Unable to move or speak, detached, "frozen" Overwhelming, inescapable Collapse (less established) Falling heart rate and blood pressure, loss of muscle tone Faintness, numbness, "switching off" The clinical importance of this sequence is difficult to overstate. A great many trauma survivors carry intense shame about what they did, or failed to do, during the event. The soldier who could not move when his friend was hit, the woman who did not scream or fight during an assault, the child who went limp while being beaten: each tends to interpret that immobility as cowardice, complicity or consent. Understanding that immobility is an automatic defensive response, not a decision, is often the first therapeutic intervention, and one of the most powerful. The evidence here is not merely theoretical. A study published in 2017 by Anna Möller and colleagues assessed nearly three hundred women who attended an emergency clinic in Stockholm within a month of a sexual assault. About seventy percent reported significant tonic immobility during the assault, and close to half reported extreme immobility. Those who experienced it were substantially more likely to develop PTSD and severe depression over the following six months. The finding has implications well beyond the clinic. It bears on how police, courts and juries interpret the absence of resistance, and on the questions clinicians ask when taking a history. Why most people recover Defensive responses are meant to be temporary. After a frightening event, most people experience some combination of intrusive memories, poor sleep, jumpiness and a reluctance to go near reminders, and for most of them these reactions fade over weeks to months. Prospective studies following people after disasters, assaults and serious accidents consistently find that early symptoms are common and that the majority of those who have them recover without specialized treatment. The question that should interest a clinician is therefore not why people develop symptoms after trauma, since almost everyone does, but why a minority fail to recover. Epidemiology puts the problem in perspective. The World Mental Health Surveys, coordinated by the World Health Organization across more than two dozen countries, found that roughly seventy percent of adults report at least one traumatic event during their lifetime. Lifetime PTSD prevalence across those surveys averaged a little under four percent, with considerable variation by country, and estimates from the United States have generally been higher, in the region of six to eight percent. The conditional risk, meaning the chance of developing PTSD given exposure to a particular kind of event, varies sharply. Events involving deliberate interpersonal violence, especially rape and sexual assault, carry far higher conditional risk than natural disasters or accidents. Being kidnapped, held captive or tortured carries high risk. Learning of the unexpected death of a loved one, because it is so common, accounts for a large share of PTSD cases in populations despite carrying a lower individual risk. What predicts who will fail to recover? Large meta-analyses of risk factors have identified a consistent but individually modest set. Characteristics before the trauma, such as prior psychiatric history, childhood adversity, prior trauma exposure and lower socioeconomic status, raise risk somewhat. Characteristics of the event itself, especially its severity, its interpersonal nature and whether the person believed they would die, raise it more. And factors after the event, particularly a lack of social support and the presence of ongoing life stress, have some of the strongest associations of all. Reactions during the trauma also matter. A meta-analysis by Emily Ozer and colleagues in 2003 found that peritraumatic dissociation, the experience of unreality, detachment or time distortion during the event itself, was among the stronger predictors of later PTSD, though subsequent research has suggested that persistent dissociation in the weeks afterward is a better predictor than dissociation in the moment. The neurobiological question underlying all of this is what normally allows the defensive response to switch off, and what prevents it from doing so in PTSD. The rest of this booklet is in large part an answer to that question, but the essential outline can be stated now. Recovery after trauma depends on several processes working together: the threat memory must be stored in a way that ties it to its original time and place; the brain must learn, through repeated safe encounters with reminders, that those reminders no longer predict danger; the regulatory systems in the prefrontal cortex must be able to apply that learning when reminders appear; and the stress hormone and autonomic systems must return to their resting set points. Avoidance, which is the natural impulse after a frightening event, interferes with every one of these processes, because it prevents the safe encounters through which new learning occurs. Learning from danger, and unlearning it Much of what is known about the neurobiology of PTSD comes from a laboratory procedure called fear conditioning, first described in its classical form by Ivan Pavlov and studied intensively in rodents and humans over the past half-century. In a typical experiment, a neutral stimulus such as a tone or a colored light is paired with something unpleasant such as a mild electric shock. After a few pairings, the neutral stimulus alone comes to elicit a defensive response: in a rat, freezing; in a human, increased skin conductance, a potentiated startle reflex and a report of expecting the shock. The neutral stimulus has become a conditioned stimulus, and the response it evokes is a conditioned response. The amygdala is essential to this learning. Its lateral nucleus is where information about the conditioned stimulus and the aversive event converges, and synaptic changes there store the association. The central nucleus then drives the defensive output. Animals with damage to the amygdala fail to acquire conditioned fear, and humans with rare bilateral amygdala damage, most famously a patient known in the research literature by the initials S.M., show strikingly reduced fear responses in both laboratory and real-world situations. The second half of the story is extinction. If the conditioned stimulus is presented repeatedly without the shock, the defensive response gradually diminishes. For a long time it was assumed that extinction erased the original memory. It does not. Extinction is new learning, a second memory that says "in this context, the tone is safe," which competes with the original threat memory. The evidence for this comes from several phenomena that every exposure therapist will recognize. After extinction, the fear response often returns spontaneously with the passage of time. It returns if the stimulus is encountered in a different context from the one in which extinction occurred, a phenomenon called renewal. And it returns if the animal or person experiences an unrelated stressor, which is called reinstatement. This finding has shaped thinking about PTSD profoundly. If extinction is a separate memory that must win a competition with the threat memory, then the critical questions become how strongly the extinction memory is formed, how broadly it generalizes, and whether the brain can retrieve it when it is needed. A body of research, much of it led by Mohammed Milad, Roger Pitman and colleagues, has shown that people with PTSD can often learn extinction in the laboratory, but they are impaired at recalling it the next day. They behave as though the safety learning had not happened. Imaging studies show that this deficit is associated with reduced activity in the ventromedial prefrontal cortex and hippocampus during the recall test, and with greater activity in the dorsal anterior cingulate cortex, a region associated with expressing fear. Two further features of threat learning help explain the clinical picture. The first is generalization. A frightening experience teaches the brain not only that the specific stimulus is dangerous but that things resembling it may be dangerous too. Some generalization is adaptive, since a predator need not look identical each time. But in PTSD generalization tends to be excessive. The veteran who was ambushed near a roadside becomes alarmed not only by roadside debris but by crowds, by heat, by the smell of diesel, by any situation in which he cannot see what is behind him. Laboratory studies confirm that people with PTSD show broader generalization gradients, responding to stimuli that only loosely resemble the original threat. The hippocampus, which supports the fine-grained distinction between similar experiences, a function called pattern separation, is thought to be central to this problem. The second feature is safety signal learning. Healthy organisms learn not only what predicts danger but what predicts its absence, and safety signals actively suppress defensive responses. People with PTSD show impaired ability to use safety cues to inhibit fear, responding to a "safe" stimulus almost as strongly as to a "dangerous" one. Clinically, this appears as the inability to feel safe even in objectively safe settings: at home with the door locked, in the therapist's office, with a trusted partner. From normal response to disorder The value of this framework for a clinician lies in its continuity. PTSD is not a strange or alien process imposed on a normal brain. It is the persistence of a set of normal defensive and learning processes after their usefulness has passed, coupled with a failure of the processes that ordinarily allow recovery. This has several practical consequences. It makes the symptoms intelligible to patients. Someone who has felt broken or crazy for years often responds with visible relief to the explanation that their nervous system learned, during the trauma, that the world was lethally dangerous, and has been acting on that learning ever since. The symptoms become evidence of a system working too hard, not evidence of weakness. It also explains why avoidance is so corrosive. Every time a person avoids a reminder, they prevent the brain from encountering the reminder safely and building the extinction memory that would allow the threat response to stand down. Avoidance is negatively reinforced, because it brings immediate relief from anxiety, and so it tends to expand over time. A patient who initially avoided only the street where the assault happened may, years later, rarely leave home. It points toward the mechanisms of effective treatment. Exposure-based therapies are, in effect, structured extinction training. Cognitive therapies alter the meanings attached to the trauma and its reminders, which shape how threatening those reminders appear. And as later chapters will show, EMDR and related approaches appear to act on the threat memory itself, changing its emotional intensity and vividness when it is recalled. Finally, the framework sets limits on what a clinician should expect. Because extinction memories compete with, rather than erase, threat memories, relapse after successful treatment is always possible, especially under stress or in new contexts. A patient who has recovered from PTSD and then encounters a new traumatic event, a major loss or a period of severe strain may find old symptoms returning. That is not a sign that the treatment failed. It is a predictable property of how threat learning works, and patients are better served by being told about it in advance, along with a plan for what to do if it happens. The two chapters that follow examine the two sides of the competition in more detail: the threat system centered on the amygdala, which fires too readily in PTSD, and the regulatory system centered on the prefrontal cortex, which fails to keep it in check. Chapter 2. A Threat System That Will Not Stand Down: The Amygdala in PTSD The amygdala has become the most famous structure in trauma neuroscience, a shorthand used in therapy rooms, self-help books and policy documents. The popularity is not unearned. Among the brain regions studied in PTSD, the amygdala has some of the most consistent findings. But the popular version of the story, in which a single structure acts as an overactive alarm, compresses a more interesting reality. The amygdala is a hub in a network, its activity in PTSD is context-dependent, and its contribution to symptoms depends heavily on the regions that feed it information and the regions that regulate its output. What the amygdala does The amygdala is not a single structure but a collection of about a dozen nuclei, each with different connections. For clinical purposes, two groupings matter most. The basolateral complex, which includes the lateral and basal nuclei, receives information from every sensory system and from the hippocampus and prefrontal cortex, and it is where associations between stimuli and their consequences are learned and stored. The central nucleus, which receives input from the basolateral complex, sends projections to the hypothalamus and brainstem that trigger the physiological and behavioral components of defense: changes in heart rate and blood pressure, release of stress hormones, freezing, startle potentiation. A third region, often grouped with the amygdala as part of what anatomists call the extended amygdala, is the bed nucleus of the stria terminalis, a structure that appears to be particularly important for sustained anxiety in response to threats that are uncertain or distant in time rather than immediate. The amygdala's job is broader than threat detection. It responds to a wide range of stimuli that are biologically significant, including rewarding ones, novel ones and socially important ones such as faces. A useful way of summarizing its role is to say that the amygdala evaluates relevance: it flags what in the environment matters for survival and wellbeing and adjusts attention, memory and physiology accordingly. Threats are the most urgent class of relevant stimuli, so they produce the most dramatic amygdala responses, but they are not the only class. One feature that has drawn attention is the speed of amygdala processing. In the 1990s LeDoux described a "low road" from the sensory thalamus directly to the amygdala in rats, which carries crude information quickly enough to trigger a response before the cortex has fully identified the stimulus, alongside a slower "high road" through the cortex. Human studies have shown that the amygdala responds to fearful faces presented too briefly to be consciously seen, and to threatening images presented to the blind half of the visual field in patients with damage to the visual cortex. The extent to which a separate subcortical pathway accounts for these effects in humans is still debated, but the general point holds: threat-relevant information can drive the amygdala and its physiological outputs without, or before, conscious recognition. This is part of the reason trauma survivors so often describe reacting to something before they know what it is. The body has braced, the heart has jumped, and only afterward does the person notice that a stranger in the street was wearing the same jacket as the attacker. What imaging studies show The first functional imaging studies of PTSD appeared in the mid-1990s. Scott Rauch, Bessel van der Kolk and colleagues used positron emission tomography to study patients listening to scripts describing their own traumatic experiences, and found increased activity in the right amygdala and associated limbic regions, along with reduced activity in Broca's area, a region involved in speech production. That second finding was later popularized as evidence that trauma is "speechless," though the result has not been consistently replicated and should not bear the weight it has sometimes been given. Many studies since have used functional magnetic resonance imaging to examine amygdala responses in PTSD. The most influential paradigms present emotional faces, trauma-related images or personalized trauma scripts, or use fear conditioning procedures. Meta-analyses pooling these studies have generally found greater amygdala activation in people with PTSD than in trauma-exposed people without the disorder, particularly in response to negative emotional stimuli. A meta-analysis by Amit Etkin and Tor Wager in 2007, comparing PTSD, social anxiety disorder and specific phobia, found that all three disorders showed amygdala and insula hyperactivity in response to negative emotional stimuli, whereas PTSD was distinguished by additional reductions in activity in medial prefrontal regions. That pattern, an overactive threat and salience system together with underactive regulation, became the core of the dominant neurocircuitry model. The amygdala responds more in PTSD not only to trauma-specific reminders but to generic signs of threat. Fearful faces, which signal that another person has detected danger, produce exaggerated amygdala responses in PTSD even when they are presented so briefly that the person is not aware of having seen them. This is consistent with the clinical picture of hypervigilance: a nervous system biased toward detecting danger in ambiguous social signals. Several important qualifications must be added. First, not every study finds amygdala hyperactivity, and effect sizes vary considerably depending on the task, the population studied and the comparison group. Second, as discussed in Chapter 5, the dissociative subtype of PTSD shows a different pattern, with reduced rather than increased amygdala activity in response to trauma reminders. Pooling patients with and without dissociative symptoms may obscure real effects in both groups. Third, the large-scale studies that have emerged over the past decade, notably through the PTSD working group of the ENIGMA consortium, which pools brain imaging data from dozens of sites around the world, have shown that structural and functional differences between people with and without PTSD are real but modest on average, with substantial overlap between groups. These are population-level findings. They are not, and may never be, diagnostic tools for individual patients. Cause or consequence? An obvious question is whether amygdala hyperactivity is caused by trauma and PTSD, or whether it is a pre-existing trait that makes some people more vulnerable. The answer appears to be both. Evidence for pre-existing vulnerability comes from prospective studies. Research on soldiers assessed before deployment, for example, has found that higher amygdala reactivity before exposure predicts greater stress symptoms afterward. A widely cited study by Admon and colleagues in Israel followed soldiers through combat training and found that pre-exposure amygdala responsiveness predicted the development of stress symptoms, while changes in hippocampal function emerged afterward. Studies of identical twins, discussed further below, point in a similar direction for some measures. Evidence for trauma as a cause comes from studies showing that exposure to severe stress alters amygdala function in previously healthy people, and from animal research. In rodents, chronic stress causes dendrites in the basolateral amygdala to grow and branch, increasing the number of synaptic connections, while dendrites in the hippocampus and medial prefrontal cortex retract. Bruce McEwen and Sumantra Chattarji, among others, documented these opposing effects, which persist for weeks after the stress has ended. If something comparable occurs in humans, it would mean that severe stress strengthens precisely the region that generates defensive responses while weakening the regions that contextualize and regulate them. The most elegant attempt to separate predisposition from consequence comes from twin studies. Roger Pitman, Mark Gilbertson and colleagues studied pairs of identical twins in which one had served in combat in Vietnam and the other had not. Among the combat-exposed twins, some had developed PTSD and some had not. If a brain characteristic was found in the combat veterans with PTSD and also in their unexposed co-twins, it was likely a pre-existing vulnerability; if it was found only in the veterans with PTSD, it was likely a consequence. Their 2002 study in Nature Neuroscience found that smaller hippocampal volume was shared by veterans with PTSD and their unexposed twins, suggesting it was a risk factor rather than a result of trauma. Later work from the same group found that heightened dorsal anterior cingulate activity also appeared to be a familial vulnerability. Other findings, such as certain physiological responses to trauma reminders, appeared to be acquired. The twin research shows that PTSD involves an interaction: some brains begin with features that make recovery harder, and trauma then produces further changes. How amygdala hyperactivity produces symptoms The symptoms of PTSD are organized in the American Psychiatric Association's DSM-5-TR into four clusters: intrusion, avoidance, negative alterations in thought and mood, and alterations in arousal and reactivity. An overactive threat system can account for a substantial portion of the first, second and fourth clusters. Intrusion symptoms include unwanted memories, nightmares, flashbacks and intense distress or physiological reactions to reminders. A threat system that responds strongly to cues associated with the trauma, and that generalizes those associations broadly, would produce exactly these phenomena. The physiological reactivity is measurable. In script-driven imagery studies, which have been conducted since the 1980s, patients listen to recordings of their own trauma narrative while heart rate, skin conductance and facial muscle tension are recorded. People with PTSD show markedly larger physiological responses to their own trauma scripts than do trauma-exposed people without PTSD, and these responses have been used as an objective measure of the disorder in research, though not in routine clinical diagnosis. Avoidance follows naturally from strong conditioned responses. If reminders produce intense distress, the person learns to avoid them, and the relief of avoidance reinforces the behavior. Arousal symptoms, including hypervigilance, exaggerated startle, irritability, poor concentration and disturbed sleep, reflect a threat system that is chronically primed. The startle reflex is a particularly useful window. It is a rapid, involuntary defensive response to a sudden stimulus such as a loud noise, and its magnitude is modulated by the amygdala and the bed nucleus of the stria terminalis. People with PTSD consistently show larger startle responses than controls in many studies, especially in contexts where some threat is anticipated, and they are less able to suppress startle in the presence of safety signals. This laboratory finding corresponds precisely to the clinical complaint of being unable to tolerate sudden noises. The fourth cluster, negative alterations in cognition and mood, is less directly explained by amygdala hyperactivity. It includes persistent negative beliefs about oneself and the world, distorted blame, persistent negative emotions such as fear, horror, anger, guilt or shame, loss of interest, detachment from others and inability to experience positive emotions. Some of these, such as persistent fear, reflect an overactive threat system. Others, such as emotional numbing and detachment, may reflect different mechanisms, including the dissociative processes examined in Chapter 5 and changes in reward processing. Research has shown that people with PTSD often have blunted responses to rewards, both behaviorally and in reward-related brain regions such as the ventral striatum, which helps explain the loss of pleasure and motivation that patients describe. The amygdala in its network The amygdala does not act alone, and the best current models place it within larger networks. Two of these deserve mention. The first is the salience network, which includes the anterior insula and the dorsal anterior cingulate cortex, and which is thought to detect events that are important to the organism and redirect attention and resources toward them. The insula plays a central role in interoception, the sensing of internal bodily states such as heartbeat, breathing and visceral sensations. People with PTSD often show increased insula activity in response to emotional stimuli, and a hyperactive salience network may explain why trauma survivors feel overwhelmed by stimuli that others barely notice, and why internal bodily sensations, such as a rapid heartbeat during exercise, can themselves trigger alarm. The second is the default mode network, a set of regions including the medial prefrontal cortex, posterior cingulate cortex and parts of the parietal and temporal lobes that are active when attention is directed inward: during self-reflection, autobiographical memory and imagining the future. Several studies have found reduced connectivity within the default mode network in PTSD and abnormal interactions between the default mode and salience networks. The psychiatrist Vinod Menon and colleagues have proposed a "triple network" model in which psychiatric disorders reflect imbalances among the salience network, the default mode network and the central executive network that supports goal-directed thinking. In PTSD, an overactive salience network may disrupt both the capacity for calm self-reflection and the capacity for focused attention on tasks, which matches patients' reports of being unable to concentrate and unable to rest. Ruth Lanius and her colleagues in Canada have extended this work to propose that trauma disrupts the brain's networks for the sense of self, particularly in people who experienced early and repeated trauma. They describe patients who struggle to experience a coherent, continuous self, who cannot easily reflect on their own thoughts and feelings, and whose sense of their own body is altered. These are among the features that distinguish complex PTSD, discussed in Chapter 7, and they suggest that for some patients the problem extends well beyond an overactive threat system. What this means in the consulting room For a clinician, the neuroscience of the amygdala has several practical implications. The first concerns how symptoms are explained to patients. A simple, accurate account, in which a threat-detection system learned during the trauma that certain cues signal danger and has been responding to those cues ever since, gives patients a framework that reduces shame and makes treatment rationale intelligible. It is important that the explanation be accurate and not overstated. Telling patients that their "amygdala is broken" or that their "thinking brain is offline" may increase hopelessness and imply that talking therapies cannot help, which is the opposite of what the evidence shows. The second concerns the therapeutic environment. A threat system that is primed to detect danger in ambiguous social signals will scan the clinician's face, tone and behavior for signs of threat. Unpredictability, sudden movements, closed doors, unexplained procedures and a clinician who appears rushed or irritated will all register. Much of what has become known as trauma-informed care, examined in Chapter 12, amounts to reducing the load on an overactive threat system so that the patient can engage in treatment. The third concerns the timing and pace of treatment. An amygdala-driven defensive response that is too intense interferes with new learning. If a patient becomes extremely distressed during exposure or processing, the session may fail to produce extinction learning; if the patient dissociates, as Chapter 5 explains, learning may not occur at all. Effective trauma therapy works within a window in which the threat memory is activated enough to be accessible to change, but not so overwhelmingly that the person cannot remain present. The phrase "window of tolerance," coined by the psychiatrist Daniel Siegel, is often used for this zone. It is a clinical heuristic rather than a precisely defined scientific construct, but it captures something that experienced therapists recognize. The fourth concerns the patient's experience of safety. Because PTSD involves impaired safety learning, a patient may intellectually know they are safe while their body continues to react as though they are not. Telling such a patient that they are safe is rarely enough. What changes the threat response is repeated, embodied experience of reminders without the anticipated catastrophe, which is why the active treatments for PTSD all involve, in some form, deliberate contact with the trauma memory. The amygdala, in short, is half of the story. The other half concerns why the regulatory systems that ought to restrain it fail to do so, and that is the subject of the next chapter. Chapter 3. When Regulation Fails: The Prefrontal Cortex and Hippocampus If the amygdala is where threat responses are generated, the prefrontal cortex and hippocampus are central to deciding when those responses are appropriate. Together they supply the context, the memory of safety and the capacity for deliberate control that allow a person who has been through something terrible to walk past a reminder without being transported back. In PTSD, this regulatory machinery is weakened, both chronically and, more dramatically, in the moment of acute stress. Understanding how and why it fails explains a great deal about the disorder, including why patients so often know their reactions are irrational and are unable to stop them. The regulatory regions The prefrontal cortex occupies the front of the frontal lobes and is the region that has expanded most in human evolution. It is not a single unit. Different sectors have different connections and functions, and for the study of trauma three are especially relevant. The ventromedial prefrontal cortex lies on the lower inner surface of the frontal lobes, just above the eyes. It has dense, reciprocal connections with the amygdala, hippocampus and hypothalamus, and it plays a key role in the regulation of emotion, the evaluation of the personal significance of events, and in particular the recall of extinction learning. In rodents, the equivalent region, the infralimbic cortex, is necessary for extinction memories to be retrieved; stimulating it reduces conditioned fear, and damaging it prevents extinction recall. In humans, the thickness of this region and its activity during extinction recall predict how well people suppress conditioned fear the day after learning that a stimulus is safe. The dorsal anterior cingulate cortex, just above and behind the ventromedial region, is associated with the expression rather than the suppression of fear. It tends to be more active when people display conditioned fear and during anticipatory anxiety. In rodents the corresponding region, the prelimbic cortex, promotes fear expression. Thus the medial prefrontal cortex contains two systems that pull in opposite directions: one promoting fear, one suppressing it. The dorsolateral prefrontal cortex, on the outer upper surface, supports working memory, planning, attentional control and the deliberate reappraisal of emotional situations. When a person tries to think differently about something distressing, for example by reminding themselves that a situation is less dangerous than it appears, the dorsolateral and ventrolateral prefrontal regions are recruited, and they appear to influence the amygdala indirectly through the ventromedial region. The hippocampus, a curved structure in the medial temporal lobe adjacent to the amygdala, is essential for forming memories of events and for representing context: where and when something happened, and what the surroundings were like. In the fear conditioning literature, the hippocampus is the structure that allows an animal to learn that a tone is dangerous in one cage and safe in another. It is also rich in receptors for cortisol and, as described in Chapter 1, participates in shutting off the stress response. Evidence of prefrontal and hippocampal dysfunction in PTSD The neurocircuitry model of PTSD, developed in its most influential form by Scott Rauch, Lisa Shin and Elizabeth Phelps in a 2006 review, proposes that the disorder involves hyperresponsivity of the amygdala together with hyporesponsivity of the ventromedial prefrontal cortex and impaired hippocampal function. Each component of this model has accumulated supporting evidence. Reduced ventromedial prefrontal activity has been reported in many studies of PTSD, particularly during exposure to trauma reminders and during extinction recall. Functional connectivity studies, which measure how closely activity in two regions rises and falls together, have frequently found weaker coupling between the ventromedial prefrontal cortex and the amygdala in PTSD, suggesting that the regulatory pathway is less effective. Structural studies have reported reduced volume or thickness in parts of the medial prefrontal cortex, including the anterior cingulate. Hippocampal findings are among the most replicated structural results in PTSD. Smaller hippocampal volume has been reported in many studies, and a large analysis from the ENIGMA PTSD consortium, published by Mark Logue and colleagues in 2018 and pooling data from about 1,800 participants across sixteen sites, confirmed a small but significant reduction in hippocampal volume in people with current PTSD compared with trauma-exposed controls. As noted in the previous chapter, twin research suggests that at least some of this difference predates trauma and represents vulnerability rather than damage. Animal studies show that chronic stress and elevated glucocorticoids can cause dendritic retraction in the hippocampus and suppress the generation of new neurons in one of its subregions, the dentate gyrus, so a contribution from stress itself is also plausible. Functionally, people with PTSD show impaired performance on some tests of hippocampus-dependent memory, particularly the ability to discriminate between similar experiences. And in imaging studies of extinction recall, the hippocampus, like the ventromedial prefrontal cortex, tends to be less engaged in people with PTSD than in those who recovered after trauma. The clinical meaning of these findings can be summarized simply. A person with PTSD may know, in a general sense, that they are safe now and that the trauma is over. But the systems that turn that knowledge into automatic, effortless suppression of defensive responses, the ventromedial prefrontal cortex retrieving the memory of safety and the hippocampus signaling that this context is not the original one, are working less effectively. The threat memory wins the competition with the safety memory more often than it should. Stress takes the prefrontal cortex offline There is a second way in which prefrontal regulation fails, and it is more dramatic. Acute stress itself temporarily impairs prefrontal function. This is not specific to PTSD. It happens to everyone, and it is part of the normal design of the defensive system. But it interacts with PTSD in ways that matter enormously for clinical care. The neurobiologist Amy Arnsten at Yale has spent decades documenting the mechanisms. The prefrontal cortex relies on networks of neurons that maintain information through persistent firing, supported by a delicate balance of the neurotransmitters noradrenaline and dopamine. At moderate levels, these neurotransmitters strengthen prefrontal network activity. Under stress, however, levels rise sharply, and at high levels they engage different receptor types and intracellular signaling pathways that open ion channels and weaken the network connections. Within minutes, the prefrontal cortex's capacity for working memory, flexible reasoning and inhibitory control declines. At the same time, high levels of noradrenaline and cortisol strengthen the function of the amygdala and of the basal ganglia, which support habitual, well-rehearsed behavior. The result is a rapid switch in how the brain controls behavior: from slow, deliberate, reflective control by the prefrontal cortex to fast, reflexive and habitual control by more primitive circuits. In genuine danger, this switch is adaptive. A person who encounters a predator, or an oncoming car, does not benefit from lengthy deliberation. But the switch has costs. The capacity to weigh alternatives, recall relevant knowledge, inhibit inappropriate responses and keep track of what is actually happening all decline. For someone with PTSD, this mechanism creates a vicious circle. Trauma reminders trigger a strong stress response; the stress response weakens the prefrontal cortex; the weakened prefrontal cortex is less able to retrieve safety learning and regulate the amygdala; and the unregulated amygdala drives an even stronger stress response. In the grip of this cycle, a patient may know intellectually that they are in a supermarket, not a war zone, but be unable to use that knowledge. The "knowing" has become inaccessible. This explains several clinical observations that otherwise seem puzzling. It explains why patients so often say that they understand their reactions are irrational but cannot stop them. It explains why coping strategies that work well in a calm session can fail completely in a real-world trigger situation. It explains the poor concentration and memory problems that are part of the diagnostic criteria. It explains impulsive anger and aggression: a person whose inhibitory control has been degraded by stress, and whose threat system is primed, is more likely to lash out. It also has implications for how institutions treat people with trauma histories. A patient who becomes agitated in an emergency department, a prisoner who reacts violently to a pat-down search, a pupil who shouts at a teacher, may each be operating in a state where the capacity for deliberate behavioral control is substantially reduced. That does not remove responsibility, but it changes what interventions are likely to work. Demanding that someone in this state "calm down and think" asks them to use precisely the faculty that stress has disabled. Prefrontal regulation and the limits of insight A popular inference from this research is that because trauma impairs the prefrontal cortex, therapies that work through language and reasoning cannot help, and only "bottom-up" body-based approaches can reach the problem. This conclusion is not supported by the evidence, and it is worth explaining why. First, the prefrontal impairment in PTSD is relative, not absolute. Outside of acute trigger states, most people with PTSD can reason, reflect and learn perfectly well, which is why they can engage in therapy at all. The goal of treatment is not to make the prefrontal cortex function under conditions of maximal stress, but to change the threat memory and strengthen regulatory pathways so that triggers produce smaller responses in the first place. Second, the therapies with the strongest evidence for PTSD, including cognitive processing therapy, prolonged exposure, cognitive therapy for PTSD and EMDR, all involve language, cognition and deliberate attention, and they produce large effects in randomized trials. If the prefrontal cortex were genuinely unable to influence trauma responses, these therapies would not work. Third, imaging studies conducted before and after successful treatment show that the regulatory regions recover. Several studies have found that successful cognitive behavioral therapy for PTSD is associated with increased activity in the ventromedial and rostral anterior cingulate cortex and decreased amygdala activity in response to trauma-related stimuli. Some studies have found increases in hippocampal volume after successful treatment, though these results are less consistent. The pattern suggests that effective therapy works at least in part by strengthening the regulatory circuitry. Pretreatment brain activity can also predict response: a study by Amit Etkin and colleagues found that people with PTSD who showed less activity in certain prefrontal regions during an emotion regulation task, and greater amygdala activity, were less likely to respond to prolonged exposure therapy. The more accurate lesson from the prefrontal research is about timing and state. Language-based and cognitive work proceeds most effectively when the patient is activated enough to engage with the trauma material but not so overwhelmed that the prefrontal cortex is effectively offline. That is precisely what experienced trauma therapists try to achieve, whichever method they use. Body-based techniques for down-regulating arousal, such as slowed breathing, grounding and orienting to the present environment, have a legitimate role in helping patients stay within that zone, but they are adjuncts to memory processing rather than replacements for it. Sleep: where regulation is restored One further aspect of prefrontal and hippocampal function deserves attention, because it is so often neglected in clinical care. Sleep is when much emotional regulation and memory consolidation takes place, and sleep disturbance is close to universal in PTSD. During rapid eye movement (REM) sleep, the brain reactivates recent emotional memories while noradrenaline levels in the forebrain fall to their lowest point of the twenty-four-hour cycle, a state in which the locus coeruleus is nearly silent. Matthew Walker and Els van der Helm have proposed that this neurochemical environment allows emotional memories to be reprocessed and integrated while their emotional charge is reduced, a process sometimes summarized as "sleep to forget, sleep to remember": the content of the memory is retained but its affective intensity is softened. Studies have found that a night of REM-rich sleep reduces amygdala reactivity to emotional images seen the day before, and that sleep deprivation increases amygdala reactivity and weakens its functional coupling with the medial prefrontal cortex. In PTSD, this process appears to break down. Noradrenergic activity during sleep is thought to remain elevated, REM sleep is fragmented, and nightmares, which in PTSD often replay the trauma with little transformation, may reflect failed rather than successful emotional processing. Sleep disturbance in the weeks after trauma predicts the later development of PTSD, and sleep problems often persist even after other symptoms have improved with treatment. This has clinical consequences. The alpha-1 adrenergic receptor blocker prazosin, which reduces the effect of noradrenaline in the brain, was widely prescribed for PTSD nightmares after early positive trials. A large trial in American veterans published by Murray Raskind and colleagues in 2018 found no significant benefit over placebo, which complicated the picture considerably, though some clinicians and guidelines continue to support its use in selected patients. Cognitive behavioral therapy for insomnia and imagery rehearsal therapy, in which patients rewrite and rehearse a changed version of a recurrent nightmare while awake, have evidence of benefit for sleep and nightmares respectively, and they are underused. More broadly, the regulatory role of sleep means that any trauma treatment plan which ignores sleep is incomplete. Bringing the two halves together The picture that emerges from Chapters 2 and 3 can be stated in a few sentences. PTSD involves a threat-detection system, centered on the amygdala and the salience network, that responds too strongly and too broadly to cues associated with danger. It involves a regulatory system, centered on the ventromedial prefrontal cortex and hippocampus, that is less effective at signaling safety and context. And it involves a stress response that, once triggered, further weakens regulation and strengthens threat responding, producing self-reinforcing cycles. Some of these features predate trauma and confer vulnerability; others are produced by trauma and its aftermath. None of them is diagnostic in individuals, but together they account for the core phenomena of the disorder. There is an important omission in this picture so far, and it concerns the trauma memory itself. Why do trauma memories intrude so vividly and so involuntarily? Why do they feel as though they are happening now rather than in the past? Why are they often fragmented, with sensory details sharp and narrative sequence missing? And how can such memories be changed? Those questions are the subject of the next chapter. Hashtags: #NeurobiologyOfTrauma #PTSD #Dissociation #SomaticTherapies #ThreatLearning #ThreatRegulation #Amygdala #PrefrontalCortex #Hippocampus #HPAAxis #AutonomicNervousSystem #FearConditioning #ExtinctionLearning #ExtinctionRecall #ThreatGeneralization #SafetyLearning #TonicImmobility #PeritraumaticDissociation #DissociativeSubtype #Interoception #SalienceNetwork #ComplexPTSD #EMDR #TraumaFocusedTherapy #FutureOfTraumaTreatment

Download the Book (PDF): Introduction In 1883 a railway engineer named Beauchamp Tower was running a set of experiments for the Institution of Mechanical Engineers. His job was mundane: measure the friction of a railway axle journal running in a partial bronze bearing, under different loads and with different methods of supplying oil. In one arrangement the journal turned with its lower half dipping in an oil bath. Tower had drilled an oil hole in the top of the bearing cap, intending to feed lubricant, and found that when the journal turned, oil rose up through the hole and spilled out. He plugged it with a cork. The cork was forced out. He plugged it with a wooden stopper, and that was forced out too. When he fitted a pressure gauge to the hole, it read a pressure considerably higher than the average pressure the load was putting on the bearing. Tower had stumbled onto something that the engineers of his day had no explanation for. The oil was not merely wetting the surfaces and making them slippery. It was carrying the load. A film of fluid, a few hundredths of a millimetre thick, was generating enough pressure on its own to hold a heavily loaded steel shaft clear of the bronze it was running in. Three years later Osborne Reynolds, then a professor at Owens College in Manchester, published a paper in the Philosophical Transactions of the Royal Society that explained Tower's results from first principles. The equation at its heart, now called the Reynolds equation, is still the foundation on which every hydrodynamic bearing in the world is designed: the bearings in car engines, in steam and gas turbines, in ship propeller shafts, in hydroelectric generators. That episode contains the idea this book is built around. Every problem of friction and wear is, at bottom, a problem about the film between two surfaces. When two solid bodies press together and slide or roll, something must separate them or they will touch, and when they touch they adhere, plough, tear, and fatigue. The separating thing might be a pressurised oil film generated by the motion itself, as in Tower's bearing. It might be a film so thin it is only a few molecules thick, laid down chemically by additives in the oil. It might be a layer of soft oxide or a deliberately applied coating of graphite, molybdenum disulphide or polymer. It might be a cloud of nanoparticles that roll, exfoliate and smear into a protective layer. Or it might be nothing at all, because the motion is too small or too slow for any film to form, and then the surfaces wear in their own peculiar ways. The discipline that studies this is called tribology, from the Greek tribos, rubbing. The word was coined in Britain in 1966, in a report for the Department of Education and Science chaired by H. Peter Jost. The Jost Report argued that British industry was losing large sums of money every year to friction, wear and their consequences, and that much of the loss could be avoided by applying what was already known. It estimated the potential saving at around £515 million a year at the prices of the day. The figure was argued over, but the point stuck: friction and wear are not minor irritations but a central cost of running a mechanical civilisation. A widely cited 2017 study by Kenneth Holmberg and Ali Erdemir estimated that roughly 23 per cent of the world's total energy consumption originates in tribological contacts, about 20 per cent spent overcoming friction and about 3 per cent spent remanufacturing worn parts and spare equipment. Whatever the precise number, the order of magnitude is not in doubt. Why the film is the right way to think There are many ways to organise a book on friction and wear. One could proceed by mechanism: adhesion, abrasion, fatigue, corrosion. One could proceed by machine element: bearings, gears, seals, brakes. One could proceed by material: metals, ceramics, polymers, coatings. Each of these is useful and each appears in what follows. But a designer facing a real machine needs a single question that cuts across all of them, and the question is this: what separates these two surfaces under these conditions, and what happens when it stops doing so? Asking that question explains a great deal. It explains why a plain journal bearing that runs for decades in a turbine at full speed can be destroyed in a few hundred starts and stops: at speed the oil film is thick and the metal never touches, while at start-up the film has not yet formed. It explains why a gear tooth can carry contact pressures of more than a gigapascal on an oil film thinner than a bacterium, and why that film depends more on the pressure-dependence of the oil's viscosity than on the viscosity itself. It explains why a wind turbine gearbox bearing that is lightly loaded and barely moving may suffer more damage than one working hard: the small oscillating motions of a parked rotor in gusty wind squeeze the lubricant out of the contact without ever pulling fresh lubricant back in. And it explains why the lubricants that work on Earth fail in space and in furnaces, where no liquid film can survive, and why engineers have turned to solid films and, increasingly, to nanoparticles engineered to build those films in place. The film framing also clarifies what designers control. They cannot change the fact that real surfaces are rough, that metals adhere, or that hard particles cut. They can choose the geometry, the speed, the clearance and the lubricant so that a fluid film forms and stays thick enough. Where that is impossible, they can choose additives, surface treatments and coatings so that a solid film forms instead. Where even that is uncertain, they can change the motion itself, or the stiffness of a joint, so that the destructive regime is avoided. Tribological design is the craft of making sure that, at every point in a machine's operating life, something suitable is in the gap. What this book covers The book moves from the bare surfaces outward. The first chapter explains why the real area of contact between two solids is a tiny fraction of the apparent area, and why that single fact underpins the classical laws of friction. The second chapter takes up friction itself: where it comes from, why the old laws of Amontons and Coulomb work as well as they do, and where they break down. The third chapter covers wear, with its handful of distinct mechanisms and the simple law, due to J. F. Archard, that engineers still use to estimate wear rates. The next four chapters build up the theory and practice of fluid films. Chapter four sets out the lubrication regimes, from full separation down to boundary films a few molecules thick, and the Stribeck curve that ties them together. Chapter five derives and interprets the Reynolds equation, showing how a converging wedge of viscous fluid generates pressure. Chapter six applies it to the journal bearing, through the classical solutions of Sommerfeld and Ocvirk, the design charts of Raimondi and Boyd, a worked example, and the thermal and stability limits that real bearings run into. Chapter seven extends the story to the concentrated contacts of gears and rolling bearings, where the film is elastohydrodynamic and the surfaces themselves deform to accommodate it. The following two chapters deal with the regime where fluid films cannot form. Chapter eight explains fretting, the damage caused by tiny oscillating motions, and the mechanics that govern it. Chapter nine applies this to wind turbine gearboxes and bearings, where fretting, false brinelling and related surface damage have become a serious reliability problem, and sets out what designers and operators can do about it. The final two chapters address environments where liquids fail altogether. Chapter ten covers the classical solid lubricants, graphite, molybdenum disulphide, polymers and soft metals, and the hard lessons of using them in spacecraft and hot machinery. Chapter eleven turns to nanoparticle lubricants, from the inorganic fullerene-like particles first studied in the 1990s to graphene, nanodiamond, two-dimensional carbides and the phenomenon of macroscale superlubricity, and asks what is proven, what is promising and what remains uncertain. How to read it The book assumes no prior knowledge of tribology. It does assume a reader comfortable with the idea of a stress, a viscosity and a differential equation, at the level of a first engineering or physics degree, though the argument can be followed without working through every equation. Where mathematics appears, it is there because the physical idea is most clearly expressed that way, and each equation is explained in words before and after. Some caution about numbers is in order. Tribology is notorious for results that depend on conditions. A friction coefficient is not a material property in the way that a density or an elastic modulus is; it depends on the pair of materials, their surface finish, the environment, the load, the speed and the history of the contact. Where the book quotes typical values, they are ranges drawn from standard handbooks and should be read as guides to the order of magnitude, not as design data. The designer who needs a number for a specific pair should test that pair under representative conditions. That is not an evasion; it is one of the central lessons of the field. Tower's corks and stoppers are a good emblem for the book as a whole. He was not looking for a new physical principle. He was measuring friction, carefully, under realistic conditions, and the surfaces told him something nobody had predicted. The history of tribology is full of such moments, and the best tribological design still begins with the humility to look closely at what is actually happening in the gap. Chapter 1. Real Surfaces and the Real Area of Contact Place a polished steel block on a polished steel plate and the two faces seem to meet everywhere. They do not. If one could shrink to the scale of a micrometre and walk across the interface, one would see a landscape of hills and valleys, and the two bodies touching only at the tips of the highest hills. Everywhere else there is a gap, sometimes tens of nanometres, sometimes several micrometres, filled with air, oil, water vapour or debris. The load that seems to be spread across the whole face of the block is in fact carried by a scattering of tiny junctions whose combined area is often less than one per cent of the face, and sometimes far less. This is the first and most important fact in tribology. Almost everything that follows, the laws of friction, the mechanisms of wear, the function of lubricants and the reason why films matter, depends on it. Roughness and how it is measured No manufacturing process produces a geometrically perfect surface. Turning leaves a helical groove from the tool tip. Grinding leaves parallel scratches from abrasive grains. Lapping and polishing leave finer, more random marks. Even a surface polished to optical quality, of the kind used for mirrors, has height variations of a nanometre or so when examined with an atomic force microscope. Cast and forged surfaces are rougher still, and surfaces that have run in service develop their own texture of wear scars and plastic flattening. Engineers describe roughness through statistical parameters measured with a stylus profilometer, which drags a fine diamond tip across the surface and records its vertical movement, or with an optical interferometer, which maps height over an area. The most common parameter is the arithmetic mean roughness, written Ra: the average absolute deviation of the profile from its mean line. The root-mean-square roughness, Rq, is the square root of the mean squared deviation and is slightly larger, by about a quarter for a roughly Gaussian surface. Typical values give a sense of scale. A turned shaft might have an Ra of 1 to 3 micrometres. A ground journal for a plain bearing is typically finished to around 0.2 to 0.4 micrometres. The raceway of a precision rolling bearing is finished to perhaps 0.05 to 0.1 micrometres. A gauge block or optical flat is better still. Ra and Rq describe how tall the hills are, but not how they are shaped or spaced. Two surfaces with the same Ra can behave quite differently in contact: one with sharp, widely spaced peaks will concentrate load on few points, while one with broad, rounded plateaus will spread it. Further parameters capture this. Skewness tells whether the profile has more peaks or more valleys; a plateau-honed engine cylinder liner, for instance, is deliberately made with negative skewness, flat bearing plateaus interrupted by deep oil-retaining valleys. Kurtosis tells how sharp the height distribution is. The slope of the asperities, and their radius of curvature at the summit, matter as much as their height when it comes to deciding whether they deform elastically or plastically. A further complication, recognised since the 1970s and studied intensively since, is that real surfaces are rough at many scales at once. Look at a ground surface at a magnification of a hundred and one sees asperities; look at the summit of one of those asperities at a magnification of ten thousand and one sees smaller asperities on top of it, and so on down towards the atomic scale. Such surfaces are often described as approximately self-affine or fractal over some range of length scales. The practical consequence is that the measured roughness depends on the instrument and its resolution, and the "real area of contact" depends on the scale at which one chooses to define contact. This multiscale view, developed in particular by Bo Persson and others in the 2000s, has refined but not overturned the older picture. From Hertz to Greenwood and Williamson The theory of how a single curved body contacts another begins with Heinrich Hertz, who in 1882, at the age of twenty-four, solved the problem of two elastic spheres pressed together. Hertz showed that the contact area is a circle whose radius grows with the cube root of the load, and that the pressure across it has a semi-elliptical distribution, highest at the centre and falling to zero at the edge. The maximum pressure is one and a half times the mean. For two steel balls the result is striking: because the contact area grows so slowly with load, the pressures are enormous. A 10-millimetre steel ball pressed onto a flat steel plate with a force of only 100 newtons produces a contact patch roughly 0.3 millimetres across and a peak pressure on the order of two gigapascals. Hertz's theory applies to smooth bodies. For rough surfaces, the decisive step was taken by J. A. Greenwood and J. B. P. Williamson in a paper published in 1966. They modelled a rough surface as a population of asperities, each with a spherical summit of the same radius, with heights distributed randomly according to some statistical law. Pressing this rough surface against a smooth flat, the tallest asperities touch first. As the load increases, those asperities deform further and their contacts grow, but, more importantly, new, shorter asperities come into contact. Each individual contact obeys Hertz's law, with area rising as load to the power two-thirds. But the population of contacts grows in number as the load increases, and for a Gaussian or exponential height distribution, the net result is that the total real contact area is very nearly proportional to the load. This was a profound result. It meant that even if every asperity behaves elastically, the real area of contact behaves as though the interface were yielding plastically at a fixed pressure: double the load and the real area doubles, while the average pressure on the real contacts stays roughly constant. That constant pressure is not the hardness of the material but something related to the elastic modulus and the slope of the asperities. Greenwood and Williamson also introduced a dimensionless number, the plasticity index, which tells whether the asperity contacts will be predominantly elastic or plastic. It combines the ratio of the composite elastic modulus to the hardness with a measure of the roughness, the square root of the ratio of the standard deviation of asperity heights to the asperity summit radius. When the plasticity index is below about 0.6, contacts are almost all elastic; when it is above about 1, plastic flow dominates at practically any load. The index depends on surface topography far more than on load, which is a useful design lesson: the way to ensure elastic contact is to make the surfaces smoother and the asperities gentler, not merely to reduce the load. Most engineering metal surfaces as machined fall in the plastic range. Running-in, the gentle initial period during which a new machine is operated at reduced load, works in large part by flattening the tallest asperities until the plasticity index falls and subsequent contact is elastic. The fully plastic limit and Bowden and Tabor Before Greenwood and Williamson, F. P. Bowden and David Tabor at Cambridge had developed a simpler and still useful picture. If the asperities yield plastically, then each junction supports a pressure roughly equal to the indentation hardness of the softer material. The real area of contact is then simply the load divided by the hardness. The arithmetic is instructive. Take a steel block weighing 10 kilograms, so pressing down with about 100 newtons, resting on a steel plate. Mild steel has an indentation hardness of something like 1.5 to 2 gigapascals. The real area of contact is then about 100 divided by 2 × 10⁹, which is 5 × 10⁻⁸ square metres, or 0.05 square millimetres. If the block's face is 100 millimetres square, its apparent area is 10,000 square millimetres. The real contact is five thousandths of one per cent of the apparent area. That small area is distributed over perhaps hundreds or thousands of junctions, each a few micrometres across. Bowden and Tabor confirmed this picture experimentally by measuring the electrical conductance across metal interfaces, which depends on the real area of contact rather than the apparent area, and found it proportional to load and independent of apparent area, as the theory predicts. Their book, The Friction and Lubrication of Solids, published in two volumes in 1950 and 1964, remains one of the foundational texts of the field. What happens inside a junction The tiny junctions where asperities meet are not simply points where two solids press together. Under the local pressure, of the order of the material's hardness, the thin surface films that normally coat metals, oxides a few nanometres thick and adsorbed water and hydrocarbons, can be broken. Where clean metal meets clean metal, atoms on either side bond. The junction becomes, in effect, a microscopic cold weld. In air the oxide films are usually tough enough to limit this, and they re-form almost instantly if broken. In a vacuum they do not re-form, and two clean metals can weld together on contact. Bowden and Tabor demonstrated this in the laboratory, and spacecraft engineers rediscovered it the hard way: the phenomenon of cold welding has caused deployment mechanisms to seize in orbit. The European Space Agency commissioned systematic studies of the problem because of it. The behaviour of a junction depends, in other words, as much on chemistry as on mechanics. The junctions also matter because almost all of the action of friction and wear takes place there. When the surfaces slide, the junctions must be sheared, or the asperities must plough through one another, or both. The energy dissipated in doing so is the frictional work. The material that is torn out of the junctions, or that fatigues and flakes off after many passes, is the wear. The heat generated at the junctions, concentrated in tiny areas for tiny times, produces so-called flash temperatures that can exceed the bulk temperature by hundreds of degrees for microseconds at a time. Those flash temperatures drive chemical reactions with the lubricant and the atmosphere, and they explain why dry sliding contacts can show oxidation and phase changes far beyond what their bulk temperature would suggest. The theory of flash temperature was developed in the 1930s and 1940s by Harmen Blok and J. C. Jaeger, and it is still used to estimate the risk of scuffing in gears. The consequence for design The fact that real contact is sparse and localised has three consequences that recur throughout this book. First, the behaviour of an interface is controlled by a very small volume of material near the surface, often the top few micrometres. That is why surface treatments and thin coatings, which change only this volume, can transform friction and wear while leaving the bulk properties of a component unchanged. A nitrided layer a few tenths of a millimetre thick, or a diamond-like carbon coating two micrometres thick, can make the difference between a component that wears out in hours and one that lasts for years. Second, anything that can get between the asperities, an oil film, an adsorbed layer of additive molecules, an oxide, a lamellar solid, a nanoparticle, acts on the very points that carry the load. A film need not be thick to be effective. A layer of fatty acid molecules a couple of nanometres thick can reduce friction between two metals by a factor of several, because it prevents the junctions from welding. Third, the separation between surfaces must be measured against their roughness. A film that is a micrometre thick is enormous for two optically polished surfaces and inadequate for two rough castings. This idea, that the ratio of film thickness to combined roughness is what really matters, will reappear when we come to lubrication regimes. It is the bridge between the topography of this chapter and the fluid mechanics of later ones. The sparse, hot, chemically active junctions at the real area of contact are where the film either does its job or fails. Understanding friction begins by asking what it takes to shear them. Chapter 2. Friction: Where the Resistance Comes From Leonardo da Vinci studied friction around the end of the fifteenth century, sliding blocks of wood across tables and measuring the force needed to move them. His notebooks record two observations that are still taught: that friction is proportional to the load pressing the surfaces together, and that it does not depend on the apparent area of contact. Leonardo did not publish, and the observations were rediscovered by the French physicist Guillaume Amontons, who presented them to the Académie Royale des Sciences in 1699. Charles-Augustin de Coulomb added a third law in 1785: that kinetic friction, once sliding is established, is nearly independent of sliding speed. He also distinguished static friction, the force needed to start sliding, from kinetic friction, the force needed to keep it going, and noted that the first is usually larger. These three statements, friction proportional to load, independent of apparent area, and roughly independent of speed, are often called Amontons' laws or the Amontons–Coulomb laws. They are summarised in the coefficient of friction, the ratio of the friction force to the normal load, written μ. For many dry metal pairs in air, μ lies between about 0.3 and 1; for lubricated metals it might be 0.05 to 0.15 in the boundary regime and 0.001 to 0.01 in a full hydrodynamic film; for polytetrafluoroethylene (PTFE) sliding slowly on steel it can be as low as 0.04 to 0.1. The laws are empirical. For over two centuries nobody could say why they held. The explanation came only with the discovery, described in the previous chapter, that the real area of contact is small and proportional to load. The adhesion model Bowden and Tabor's explanation runs as follows. The real area of contact, A, is the load W divided by a pressure that stays roughly constant, the hardness H in the plastic case or an elastic equivalent in the Greenwood–Williamson case. To slide one surface over the other, the junctions formed at the real contacts must be sheared. If the shear strength of a junction is τ, the friction force is τ multiplied by A. Combining these, the friction force is W multiplied by τ/H, and the coefficient of friction is simply τ/H. This one ratio explains all three laws. Friction is proportional to load because the real area is. It is independent of apparent area because the real area does not depend on apparent area. It is roughly independent of speed because neither shear strength nor hardness changes much with speed over ordinary ranges. It also explains why soft films on hard substrates are such effective lubricants. If a thin layer of a soft material, with low shear strength, is laid on a hard substrate, the load is supported by the hard substrate, so the real area of contact is small, but sliding takes place in the soft film, so the shear strength is low. The friction coefficient becomes the shear strength of the soft film divided by the hardness of the hard substrate, which can be very small. A thin layer of indium, lead or silver on steel, a layer of graphite or molybdenum disulphide, or a monolayer of adsorbed fatty acid all exploit this principle. Plain bearing alloys, the white metals or babbitts, tin- and lead-based alloys used since Isaac Babbitt patented them in 1839, embody a related idea: a soft, conformable matrix with harder particles dispersed in it, so that the bearing can embed dirt and conform to the shaft while still carrying load. The simple adhesion model has a problem, and Bowden and Tabor recognised it. For a pure metal, the shear strength is roughly one-sixth of the hardness, which would predict a friction coefficient of about 0.17. Clean metals in vacuum show friction coefficients much higher than that, often greater than 1 and sometimes so high that the surfaces seize. The reason is junction growth. When a junction that is already yielding under normal pressure is also subjected to a tangential force, the combined stresses cause further plastic flow, and the junction grows. The real area of contact increases as the surfaces try to slide, and in the absence of any contaminant film, it can keep growing until the two bodies are effectively welded. In air, oxide and adsorbed films limit junction growth by providing a weak plane at the interface, which is why the friction of metals in ordinary air is moderate. The difference between friction in air and in vacuum is a vivid demonstration that friction is a property of the interface, not of the bulk materials. Ploughing and the role of hardness differences Adhesion is not the only source of friction. When a hard asperity, or a hard particle trapped between the surfaces, presses into a softer surface and slides, it must push material aside. This ploughing or deformation component adds to the friction force. For a conical asperity of a given slope, the ploughing contribution to the friction coefficient depends on the slope angle: sharp asperities plough hard, blunt ones barely at all. For most engineering surfaces the asperity slopes are only a few degrees and the ploughing contribution is small, around 0.05 or less. But when hard particles, wear debris, dust or abrasive grit enter a contact, ploughing can dominate, and it is the link between friction and abrasive wear. The total friction is often treated, approximately, as the sum of adhesion and ploughing terms. This is a simplification, because the two interact: ploughing raises local stresses and affects junction growth, and adhesion at the ploughing front affects how material flows. More detailed models, by Nam Suh and others in the 1970s and 1980s, added a third component from the trapped wear particles themselves, and argued that the evolution of friction during running-in reflects changes in the balance between these components as debris builds up and the surfaces smooth. Static friction, stick-slip and the rate-and-state picture Coulomb's distinction between static and kinetic friction has practical consequences. If the static coefficient exceeds the kinetic, then a system driven through a spring, which is to say almost any real system because every machine has some compliance, can show stick-slip motion. The surfaces stick, the spring winds up, the static friction is overcome, the surfaces slip, the friction drops to the kinetic value, the spring unwinds, the motion slows, and the surfaces stick again. Stick-slip produces the squeal of brakes, the chatter of machine tool slideways, the creak of a door hinge and the note of a bowed violin string. In precision machines it is a serious problem, limiting positioning accuracy. Why should static friction exceed kinetic? Part of the answer is that the real area of contact grows with time while the surfaces are at rest, through slow creep of the junctions. The longer two surfaces sit together, the stronger they stick. Experiments by Rabinowicz in the 1950s and later by geophysicists studying rock friction showed that static friction increases roughly with the logarithm of the time of stationary contact. When surfaces slide, each junction has a lifetime equal to the asperity size divided by the sliding speed; at higher speeds junctions have less time to grow, so friction tends to fall with speed. This velocity-weakening behaviour, if strong enough, produces stick-slip instability. Geophysicists studying earthquakes, which are essentially stick-slip events on faults, formalised this in the late 1970s and early 1980s as rate-and-state friction laws, associated with James Dieterich and Andy Ruina. In these laws the friction coefficient depends on the sliding velocity and on a state variable that represents the average age, and hence strength, of the contacts. The laws describe well a wide range of materials, rocks, metals, polymers and paper among them, and they have become a standard tool for analysing stick-slip and frictional stability in engineering as well as in seismology. The design remedies for stick-slip follow from the physics. Make the static and kinetic coefficients closer, which is what slideway oils with special friction-modifying additives do. Make the system stiffer, so that the stored energy released in a slip is small. Add damping. Or use a lubricant regime in which friction rises with speed rather than falling, which is the case in a full hydrodynamic film. That last remedy, the idea that friction can be made to increase with speed, points ahead to the Stribeck curve and the lubrication regimes discussed in chapter four. Rolling friction Rolling is far easier than sliding, which is why wheels and rolling bearings exist. A steel ball bearing typically has a friction coefficient, defined as the frictional torque divided by the load and the bore radius, of perhaps 0.001 to 0.005, compared with 0.1 or more for a boundary-lubricated sliding contact. But rolling resistance is not zero, and its sources are worth knowing because they explain the behaviour of rolling element bearings and gears later in the book. There are three main sources. The first is elastic hysteresis: as a ball rolls, the material ahead of it is compressed and the material behind it is released, and no real material returns all the energy stored in compression. The loss is small for hardened steel, larger for polymers and rubber; the high rolling resistance of a car tyre is largely hysteresis in the rubber. The second is microslip within the contact. Because the contact patch has finite size and the two bodies deform differently, points within the patch cannot all roll without slipping. In a deep-groove ball bearing the ball contacts a curved groove, so different points on the contact patch are at different radii from the ball's axis of rotation and travel at different speeds; some slip forward and some backward. This Heathcote slip, named for the engineer who analysed it in the 1920s, contributes friction and, under poor lubrication, wear. The third is the viscous drag of the lubricant, both in the inlet to each contact and in the churning of oil within the bearing. At high speeds, churning losses often dominate the friction of a bearing. Friction is a systems property The most important practical lesson about friction is that the coefficient of friction is not a material constant. It depends on the pair of materials, their surface finish, the environment, the lubricant, the load, the speed, the temperature and the history of the contact. Handbooks list values, and they are useful for orientation, but they come from particular test rigs under particular conditions, and different rigs often disagree. Friction of steel on steel can vary from below 0.1 to above 1 depending on whether the surfaces are oxidised, how much moisture is present and whether they have been cleaned. Graphite, an excellent lubricant in humid air, has high friction in dry nitrogen or vacuum, for reasons explored in chapter ten. Molybdenum disulphide behaves in the opposite way. This systems view, developed explicitly by Horst Czichos in the 1970s, treats the tribological contact as a system with inputs such as load, motion and temperature, outputs such as useful motion and transmitted force, and losses such as friction and wear, and holds that friction and wear rates can only be properly specified for the whole system, not for individual materials. It is why serious tribological testing tries to reproduce the real contact geometry, motion and environment, and why results from a standard pin-on-disc test must be transferred to a real machine with great care. For the designer, the useful conclusion is that friction is controlled by what happens in the junctions, and the most powerful lever over those junctions is what lies between the surfaces. Replace metal-to-metal shear with shear in a soft film and friction falls by a factor of several. Replace it with shear in a fluid film and friction falls by a factor of a hundred. The rest of the book is about how to put those films in place, and keep them there. Chapter 3. Wear: How Surfaces Lose Material Friction costs energy; wear costs machines. Most mechanical equipment is retired not because it breaks suddenly but because its working surfaces have changed shape enough that it no longer does its job: clearances open up, gear teeth lose their profile, seals leak, bearings run rough and noisy. Ernest Rabinowicz, who taught tribology at MIT for many years, liked to point out that only a small fraction of the material in a worn-out car has actually been lost. A few grams removed from the right surfaces is enough to end its useful life. Wear is more varied than friction. Friction is, broadly, one phenomenon with a couple of contributing mechanisms. Wear is several distinct phenomena that happen to share a name, each with its own physics, its own appearance under the microscope and its own remedies. A designer who diagnoses the wrong mechanism will apply the wrong cure, and some cures for one mechanism make another worse. Hardening a surface, for instance, protects against abrasion but can make a component more prone to brittle fatigue cracking. The principal mechanisms Most classifications recognise four basic mechanisms, adhesive, abrasive, fatigue and tribochemical (or corrosive) wear, plus a group of special cases defined by their kinematics, of which fretting and erosion are the most important. Table 1 summarises them. Table 1. The principal wear mechanisms compared. Mechanism What happens Typical appearance Principal remedies Adhesive Junctions weld and tear, transferring material Transfer layers, tearing, scuffing, seizure Separating films; dissimilar or non-metallic pairs; EP additives Abrasive Hard asperities or particles cut and plough Parallel grooves and scratches Harder surface; filtration and sealing; remove debris Surface fatigue Repeated stress cycles initiate and grow cracks Pits, spalls, micropits Thicker film; cleaner steel; smoother surfaces; lower stress Tribochemical Reaction films form and are rubbed away Oxide or reaction-film debris, smooth worn surfaces Control chemistry and temperature; inhibitors; coatings Fretting Small-amplitude oscillation; debris trapped in contact Oxide debris (red-brown on steel), cracks at contact edge Eliminate slip; change stiffness; coatings; special greases Adhesive wear When asperity junctions weld and are then sheared, the shear does not always happen at the original interface. If the junction is stronger than one of the parent materials, the fracture occurs inside the weaker body, and a fragment of it is transferred to the other surface. Repeated transfer builds up lumps that eventually break loose as wear particles. In mild form, adhesive wear produces a fine transfer layer and modest wear; in severe form it produces scuffing (also called scoring or galling), in which the surfaces tear and roughen rapidly, and finally seizure, in which the junctions grow and weld until motion stops. Adhesion is most severe between clean, chemically similar metals that are mutually soluble, such as steel on steel or aluminium on aluminium. Pairs that are insoluble in one another, such as steel against silver or lead, have much lower tendency to weld, which is part of why lead-tin-copper and white-metal bearing alloys run well against steel shafts. Austenitic stainless steels and titanium alloys are notorious for galling: threads seize, valve stems lock. The cures are to ensure a separating film, to choose dissimilar materials, or to use a coating such as a nitride, a carbon-based coating or a solid lubricant. Abrasive wear Abrasive wear occurs when hard asperities or hard particles cut or plough a softer surface. It is called two-body abrasion when the hard asperities are fixed on the opposing surface, as with a file or sandpaper, and three-body abrasion when loose particles roll and slide between the surfaces. Three-body abrasion is usually milder per unit of sliding distance, because loose particles spend much of their time rolling, but it is ubiquitous: dust ingested past seals, wear debris circulating in oil, sand in the grease of an exposed bearing. The most important rule of abrasive wear is that the surface must be harder than the abrasive. When the ratio of surface hardness to abrasive hardness rises above about 1.2, the abrasive particles can no longer indent effectively and wear falls sharply. Quartz, the commonest natural abrasive, has a hardness around 11 gigapascals, harder than any ordinary hardened steel. That is why filtration and sealing are so important: no affordable steel can resist sand, but a good filter can keep sand out. The cleanliness of hydraulic oils and gearbox lubricants is specified by particle count standards such as ISO 4406, and much of the improvement in the reliability of hydraulic systems since the 1970s has come from finer filtration rather than from better materials. Surface fatigue Rolling contacts, in bearings and gears, wear in a different way. Rolling produces little sliding and therefore little adhesion or abrasion, but every time a ball passes over a point on a raceway, the material beneath that point goes through a cycle of stress. Hertz's analysis shows that the maximum shear stress lies not at the surface but below it, at a depth of roughly half the contact half-width for a cylinder and about 0.48 times the contact radius for a sphere. After millions or billions of cycles, cracks initiate, often at non-metallic inclusions in the steel near the depth of maximum shear stress, and grow until a piece of the surface breaks away, leaving a pit or spall. Once spalling begins, the debris and the disturbed surface accelerate damage, and the bearing fails. The fatigue life of rolling bearings is statistical: under identical conditions, a batch of bearings fails over a wide spread of times. The standard rating method, originated by Gustaf Lundberg and Arvid Palmgren of SKF in the late 1940s and now codified in ISO 281, defines the basic rating life L10 as the life that 90 per cent of a large group of identical bearings will reach or exceed. For ball bearings, life varies inversely with the cube of the load. Double the load and the life falls by a factor of eight. Later refinements, notably by Ioannides and Harris in the 1980s, introduced a fatigue limit stress below which the material does not fail, and life modification factors for lubrication and contamination. The single biggest factor in modern bearing life is not load but lubrication: the ratio of oil film thickness to surface roughness, discussed in chapters four and seven, can change the predicted life by an order of magnitude. Steel cleanliness has improved dramatically since Lundberg and Palmgren's time. Vacuum degassing and better steelmaking practice have reduced the number and size of oxide inclusions, and with them the incidence of subsurface-initiated fatigue. In modern bearings, failure is increasingly initiated at the surface, from dents made by debris particles, from micropitting where films are thin, or from the white etching cracks discussed in chapter nine. Micropitting deserves a separate mention. It is a form of surface fatigue on the scale of the roughness itself, producing pits a few to some tens of micrometres deep that give the surface a grey, frosted look; in gears it is called grey staining. It occurs when the oil film is thin compared with the roughness, so that individual asperities carry high local stresses as they pass through the contact. Micropitting can stabilise, or it can progress to change the gear profile and initiate full-scale pitting. It is a significant concern in the high-speed stages of wind turbine gearboxes. Tribochemical wear Surfaces react with their environment: steel oxidises, and lubricant additives react with steel to form protective films. Sliding accelerates these reactions, through flash temperatures and the constant exposure of fresh metal. If a reaction film is formed and rubbed off repeatedly, the material it consumed is lost. This is tribochemical wear. In mild oxidational wear of steel, it can be benign, producing smooth surfaces and fine oxide debris at low rates; much of the steady, mild wear of machines in service is of this kind. Sometimes the reaction film is protective and the trick is to keep forming it faster than it wears; that is exactly how antiwear additives such as zinc dialkyldithiophosphate work, as chapter four explains. When the chemistry is aggressive, for example in the presence of water, acids or sulphur at high temperature, tribochemical wear becomes corrosive wear and can be severe. Fretting, erosion and others Fretting is wear caused by very small oscillating motions between surfaces nominally at rest relative to each other, such as press fits, bolted joints, splines and the rollers of a stationary bearing subject to vibration. Its peculiarity is that debris cannot escape and fresh lubricant cannot enter, so the contact develops its own chemistry and mechanics. Fretting is so important to the later part of this book that chapter eight is devoted to it. Erosion is wear by impact of particles or liquid droplets carried in a fluid: sand in pipelines and pumps, rain on wind turbine blade leading edges, steam droplets on the last-stage blades of steam turbines. Cavitation erosion, caused by the collapse of vapour bubbles against a surface, damages pump impellers, ship propellers and, occasionally, the soft linings of heavily loaded journal bearings where the oil film ruptures and reforms. Archard's law In 1953 J. F. Archard, working at the Associated Electrical Industries research laboratory, proposed a simple equation for sliding wear. The volume of material removed, V, is proportional to the normal load W and the sliding distance s, and inversely proportional to the hardness H of the softer surface: V = K × W × s / H The constant K is a dimensionless wear coefficient. Archard derived the equation from the idea that each asperity junction has some probability of producing a wear particle of size comparable to the junction; K is then interpreted as that probability. In practice, engineers often combine K and H into a single dimensional wear coefficient k = K/H, expressed in cubic millimetres per newton-metre, which is easier to measure directly. The law's power lies in its simplicity and its range. Measured values of K span more than eight orders of magnitude, from about 10⁻² for clean metals sliding in vacuum to below 10⁻¹⁰ for well-lubricated steels in mild conditions. That range is the real message: the wear coefficient is not a property of a material but of a system, and modest changes in conditions can shift it by factors of a hundred. A transition from mild to severe wear, triggered by a rise in load, speed or temperature that breaks down a protective oxide or lubricant film, often produces such a jump. Archard's law is therefore best used in two ways. First, to scale wear within a regime: if a test establishes k for a material pair under realistic conditions, the law predicts how wear will change with load and distance, as long as the mechanism does not change. Second, as a framework for wear maps. In the 1980s Lim and Ashby at Cambridge plotted wear rates and mechanisms for steel on steel as a function of normalised pressure and sliding speed, producing maps that show where mild oxidational wear, severe adhesive wear, melt wear and seizure dominate. Such maps make the point that the most important design decision is choosing a regime, not fine-tuning a coefficient. Archard's law also underlies many wear calculations in practice. It is used to estimate the life of dry polymer bushings, where manufacturers publish a PV limit, the product of contact pressure and sliding velocity above which frictional heating makes wear catastrophic, alongside a wear factor for use within it. It is used, with care, to predict the change of profile of cam followers and gear teeth, by integrating local pressure and sliding over many cycles. And it is used, as chapter eight explains, in modified energy-based forms, to predict fretting wear. Running-in and the steady state New machines usually wear faster at first than they do later. During running-in, the tallest asperities are flattened, the surfaces conform to one another, debris is flushed away and chemical films form. Wear rate falls and friction often falls too, and the machine enters a long period of steady, mild wear. Eventually, accumulated changes, or a fatigue process, lead to accelerating wear and failure. This three-stage history, sometimes called the bathtub curve of wear, is familiar to anyone who has run a new engine gently for its first thousand kilometres. Running-in is an under-appreciated design tool. A gear set run in carefully at reduced load develops a smoother surface finish and a thicker relative oil film than one subjected to full load from the start, and will be less prone to micropitting and scuffing for its whole life. Some manufacturers specify running-in procedures, some use special running-in oils with chemically active additives, and some finish surfaces so finely, by superfinishing or isotropic chemical polishing, that running-in is effectively done in the factory. The common thread is the one that runs through this book: a mild wear regime is one in which a protective film, whether fluid, chemical or a smoothed surface, keeps the asperities from destructive interaction. Severe wear is what happens when it does not. Hashtags: #TribologyInMechanicalDesign #Tribology #Friction #Wear #HydrodynamicLubrication #ReynoldsEquation #LubricationRegimes #StribeckCurve #JournalBearings #ElastohydrodynamicLubrication #RealAreaOfContact #SurfaceRoughness #AsperityContact #GreenwoodWilliamsonModel #BowdenTaborTheory #AmontonsCoulombLaws #ArchardsWearLaw #AdhesiveWear #AbrasiveWear #SurfaceFatigue #TribochemicalWear #FrettingWear #SolidLubricants #NanoparticleLubricants #FutureOfTribologicalDesign

Download the Book (PDF): Introduction Tuberculosis has killed more human beings than any other infectious disease in history, and it is still trying. In 2024, by the World Health Organization's reckoning, about 10.7 million people fell ill with it and roughly 1.23 million died. Those figures describe a disease that antibiotics were supposed to have finished seventy years ago. The drugs exist. Most of the people who die of tuberculosis die of a curable illness, undiagnosed or half-treated, for want of a laboratory, a clinic, or an uninterrupted supply of pills. It is the paradigmatic disease of inequality: a bacterium that flourishes wherever people are poor, crowded, undernourished, or untreated for HIV, and that has largely retreated from the places rich enough to have beaten it back. This booklet is about the hardest part of that failed endgame: the strains of Mycobacterium tuberculosis that no longer respond to the drugs on which the whole edifice of tuberculosis control was built. When the bacillus becomes resistant to rifampicin and isoniazid — the two agents that make ordinary tuberculosis curable in six months — the disease becomes something else. Historically it became a sentence to eighteen months or two years of treatment with a handful of weaker, more toxic drugs, including daily injections that could leave a patient deaf; and even then, across the world, barely half of those who started finished cured. Multidrug-resistant tuberculosis was, until very recently, one of the grimmest diagnoses in medicine: expensive, brutal, and frequently fatal despite everything. The central claim of the pages that follow is that this picture has been transformed in little more than a decade, and that the transformation holds a lesson worth stating plainly. The gains did not come from a single miracle drug. They came from re-engineering the entire chain of care — the way resistance is detected, the way regimens are built, and the way treatment is delivered — around two principles that the old system violated at every step: speed and tolerability. A molecular test that reports rifampicin resistance in two hours instead of the two or three months a culture demanded. Regimens assembled from new oral drugs that cure in six months instead of twenty, without a single injection. A willingness, finally, to throw out agents that worked but maimed. Each of these advances matters on its own. Together they amount to a different disease. That is the optimistic half of the argument, and it is real. A patient diagnosed today with rifampicin-resistant tuberculosis in a reasonably equipped programme can expect a rapid diagnosis, an all-oral regimen of four drugs taken for six months, and a cure rate that trial populations have pushed close to ninety percent. A decade ago that sentence would have read as science fiction. The combination of bedaquiline, pretomanid, linezolid, and moxifloxacin — the regimen the shorthand calls BPaLM — did to resistant tuberculosis what the discovery of rifampicin once did to the drug-susceptible kind: it collapsed the treatment from an ordeal into a course. The pessimistic half is equally real and must be stated with equal force, because a booklet that only told the good news would be dishonest about the field as its practitioners actually experience it. The new tools reach only a fraction of the people who need them. Of the roughly 390,000 people who developed multidrug- or rifampicin-resistant tuberculosis in 2024, fewer than half were ever diagnosed and started on treatment; the rest were never found, or were found too late, and many of them died or transmitted their resistant strains to others before anyone knew what they had. The rapid tests that make everything else possible covered barely half of the world's notified cases. And the whole enterprise rests on a financial base that is both inadequate — global tuberculosis funding stands at roughly a quarter of the agreed target — and, as of this writing, suddenly and severely threatened by the collapse of donor commitments that had underwritten tuberculosis programmes across Africa and Asia for two decades. The WHO has modelled that abrupt funding cuts could translate into millions of additional cases and deaths over the coming decade. Progress against tuberculosis has never been self-sustaining. It is bought, year by year, and the bill has lately gone unpaid. There is, finally, a biological threat that runs underneath all of this. The new regimens work because the bacillus has not yet learned to defeat them at scale. But resistance to bedaquiline, the keystone of every modern regimen, is already being documented in clinics from South Africa to Eastern Europe, and the mechanisms by which it arises are exactly the ones that history should have taught us to expect. The story of tuberculosis chemotherapy is, from its very first chapter in 1944, a story of resistance outrunning the drugs. We have been given a short lead over the pathogen. Whether we keep it will depend less on pharmacology than on stewardship, diagnostics, and the unglamorous discipline of making sure that every patient who starts the right regimen finishes it. This book is written for clinicians who treat tuberculosis and for the public-health readers who design the programmes within which they work, though it assumes no specialist knowledge and defines its terms as it goes. It moves from the pathogen to the system. The early chapters explain what M. tuberculosis is, why it is so uniquely difficult to kill, and how its resistance is made — because none of what follows makes sense without that foundation. The middle chapters trace the two technological revolutions that reshaped the field: the molecular diagnostics, GeneXpert chief among them, that made rapid detection of rifampicin resistance possible; and the new oral drugs that made short, injection-free treatment possible. The later chapters turn to the hard practical questions — how the current regimen menu is actually chosen and managed, where the new drugs' toxicities and interactions bite, and how resistance to them is emerging — and then to the system that surrounds the drugs: the economics, the prevention, the vaccines in the pipeline, and the social conditions that decide, more than any molecule, who gets tuberculosis and who dies of it. The argument that emerges from all of this is that the resistant-tuberculosis problem has become, to an underappreciated degree, a solved problem that remains unsolved. The science has, for the moment, caught up with the pathogen. What has not caught up is the capacity to deliver the science to the people who need it, consistently and at scale, and to protect the fragile advantage the new drugs have bought. That gap — between what is possible and what is done — is where the remaining deaths live. It is also where the next decade of the fight will be won or lost. The Patience of the Bacillus To understand why drug-resistant tuberculosis is such a formidable problem, you have to begin with the organism, because almost everything that makes the disease hard to treat — resistant or not — flows from the peculiar biology of Mycobacterium tuberculosis. It is not a fast, florid pathogen like the agents of cholera or sepsis, which announce themselves within hours and are often cleared or lethal within days. It is a slow one. It has made patience into a survival strategy, and patience is exactly the quality that defeats the ordinary logic of antibiotics. Robert Koch identified the bacillus in 1882, in a lecture to the Berlin Physiological Society that is still studied as a model of experimental reasoning. He stained the organism, cultured it, reproduced the disease in animals, and recovered the same organism again — the sequence that became known as Koch's postulates. What he had found was a rod-shaped bacterium that grows extraordinarily slowly. Most pathogenic bacteria divide every twenty or thirty minutes; M. tuberculosis divides roughly once every sixteen to twenty-four hours. A colony that a fast organism would raise on a culture plate overnight takes the tubercle bacillus three to six weeks to produce. This single fact has shaped the history of tuberculosis diagnosis and treatment more than any other. It is why culture-based diagnosis has always been measured in weeks, why drug-susceptibility testing by the classical method takes months, and why treatment, which must outlast the organism's capacity to persist, is measured not in days but in months. You cannot treat a slow infection quickly. The slowness is bound up with the organism's most distinctive feature: its cell envelope. Most bacteria wrap themselves in a relatively simple membrane and cell wall. The mycobacterial envelope is a fortress. Outside the ordinary plasma membrane and a layer of peptidoglycan lies a vast, waxy coat built largely from mycolic acids — long-chain fatty acids, sixty to ninety carbon atoms in length, found nowhere else in such abundance. This mycolic-acid layer, cross-linked and nearly impermeable, is the reason the organism resists drying, survives for weeks in expelled droplets, shrugs off many disinfectants, and keeps out a large fraction of the antibiotics that would otherwise kill it. It is also the reason tuberculosis bacteria stain the way they do. The classic Ziehl-Neelsen method, which floods a smear with hot carbolfuchsin and then tries to wash the red dye out with acid alcohol, works because the mycolic coat holds the dye against the acid — hence the term "acid-fast bacilli," the phrase that still appears on every sputum-smear report. The waxy wall is simultaneously the organism's armour and its signature. The envelope is not merely a passive barrier. Its lipids are active participants in disease, modulating the immune response, helping the bacillus survive inside the very cells sent to destroy it. For M. tuberculosis is, above all, an intracellular parasite of the macrophage — the scavenging immune cell whose job is to engulf and digest invaders. When a person inhales tubercle bacilli, deposited deep in the lung in droplets small enough to reach the alveoli, the alveolar macrophages do what they are designed to do and swallow them. In most infections that would be the end of the matter. But M. tuberculosis has evolved to survive and replicate inside the macrophage, arresting the normal maturation of the phagosome — the internal vacuole in which the cell would ordinarily digest what it has eaten — and turning the immune cell into a sanctuary and a vehicle. The organism's great trick is to live inside the one cell type most equipped to kill it. What follows is a standoff that can last a lifetime. The immune system walls the infection off inside an organized structure called a granuloma — a knot of macrophages, lymphocytes, and other cells that surrounds the bacilli and contains them. In the great majority of infected people, the granuloma holds. The bacteria are not eradicated but are held in check, often for decades, in a quiescent state that clinicians call latent tuberculosis infection. Roughly a quarter of the world's population is estimated to carry M. tuberculosis in this latent form. Most will never develop disease. The lifetime risk of a latent infection reactivating into active, transmissible tuberculosis is on the order of five to ten percent — but that risk is transformed by anything that weakens cellular immunity, which is why HIV, which destroys precisely the T-cells that keep the granuloma intact, is the single most powerful driver of tuberculosis on earth. A person co-infected with HIV and latent tuberculosis may face an annual reactivation risk of ten percent rather than a lifetime risk of ten percent — a roughly tenfold-per-year amplification that turned the two epidemics, from the 1990s onward, into a single catastrophe across sub-Saharan Africa. The granuloma is also where the problem of persistence is born, and persistence is the hinge on which the entire difficulty of tuberculosis treatment turns. Inside these structures, conditions are hostile: oxygen is scarce, nutrients are limited, the pH shifts. The bacilli respond by slowing almost to a stop. They stop dividing, downshift their metabolism, and enter a dormant or near-dormant state in which they are phenotypically tolerant of drugs — not genetically resistant, but simply not doing the things that most antibiotics need them to be doing in order to be killed. A drug that attacks cell-wall synthesis cannot do much to a bacterium that has stopped building its wall. A drug that disrupts protein synthesis has little purchase on a cell that has nearly stopped making protein. These dormant persisters are the reason tuberculosis cannot be cured in the week or two that would clear most bacterial infections. The rapidly dividing bacilli in an open lung cavity die quickly under treatment — which is why patients become non-infectious and feel better within weeks — but a residuum of slow, tolerant organisms survives, and it is to kill that residuum, and so to prevent relapse, that treatment must grind on for months after the patient appears cured. Here the logic of tuberculosis chemotherapy reveals itself, and it is a logic that applies with special force to the resistant disease. Treatment has two jobs that require different tools. The first is to kill the large, actively multiplying population fast — to reduce the bacterial load, render the patient non-infectious, and prevent clinical deterioration. Isoniazid, historically, has been the great early bactericidal agent for this. The second job is sterilization: the slow killing of the persistent, semi-dormant subpopulation over months, so that when treatment stops the disease does not come back. Rifampicin, with its unusual ability to act against slowly metabolizing organisms, and pyrazinamide, which works in the acidic environment of the granuloma, are the sterilizing workhorses of standard therapy. A good regimen needs both capabilities, and it needs enough different drugs acting by enough different mechanisms that the bacillus cannot escape through any single mutation. Lose rifampicin — the most important sterilizing drug — and you lose the ability to cure quickly. That is precisely what resistance to rifampicin means, and it is why rifampicin resistance, rather than resistance to any other single drug, is the dividing line that defines the whole field. There is one more feature of the organism that bears directly on resistance, and it concerns reproduction. M. tuberculosis is, for practical purposes, a clonal organism. Unlike many bacteria, it does not readily swap genetic material with its neighbours through the horizontal transfer of plasmids — the mechanism by which resistance genes spread so alarmingly among gut bacteria and hospital pathogens. The tubercle bacillus does not pick up resistance from other bacteria. It can only acquire resistance the slow way: through spontaneous mutations in its own chromosome, arising at random as the organism divides, and then selected for when a drug kills the susceptible majority and leaves the mutant to multiply. This is a crucial and in some ways reassuring fact. It means resistance in tuberculosis is not an infectious cargo passed around a community in the way a plasmid is. But it also means that resistance, once it arises in a patient through inadequate treatment, is written permanently into that strain's chromosome and can be transmitted, whole, to the next person infected — a point we will return to, because the distinction between resistance that a patient acquires and resistance that a patient catches turns out to be central to controlling the epidemic. The slow growth, the impenetrable wall, the intracellular refuge, the lifelong latency, the dormant persisters, the clonal chromosome: these are not six separate facts about a microbe. They are a single coherent strategy, honed over the tens of thousands of years that M. tuberculosis has co-evolved with human beings, and that strategy is the reason tuberculosis treatment is long, combination treatment is mandatory, and resistance, once established, is so difficult to reverse. Everything in the chapters that follow — the two-hour molecular test that changed diagnosis, the toxic injections that defined a generation of treatment, the new oral drugs that have begun to undo that legacy — is in some sense a response to the patience of the bacillus. To beat an organism built for the long game, medicine has spent a century learning to play it too. The recent revolution, the one this book is about, consists largely of learning how to win that long game faster, and with less harm to the patient, than anyone thought possible. How Resistance Is Made Drug resistance in tuberculosis is not a mysterious emergent property. It is the predictable result of a specific and well-understood biological process, colliding with a specific and well-understood failure of treatment. Understanding both halves of that collision is the key to everything that follows, because the molecular diagnostics of the modern era work by reading the genetic marks that resistance leaves, and the whole philosophy of regimen design is an attempt to deny the bacillus the openings through which resistance arises. To see why a two-hour test can report rifampicin resistance, and why a four-drug regimen is safer than a three-drug one, you first have to see exactly how a susceptible strain becomes a resistant one. Mutation and selection Begin with the arithmetic of mutation. Every time a bacterium divides, its DNA is copied, and copying is never perfect. Random errors creep in at a low but non-zero rate. For any given gene, the chance that a particular mutation conferring resistance to a particular drug will arise spontaneously in a single bacterial division is tiny — for rifampicin, on the order of one in a hundred million cell divisions; for isoniazid, perhaps one in a million. These numbers sound like safety, and for a single drug against a small infection they would be. But tuberculosis is not a small infection. An adult with an open pulmonary cavity may carry a bacterial population of a hundred million to a billion organisms or more. In a population that large, the mutations that confer resistance to any single drug are not a remote possibility; they are a statistical certainty. At any moment, in any large tuberculous lesion, there already exist a few bacilli carrying a rifampicin-resistance mutation and a few more carrying an isoniazid-resistance mutation, generated by nothing more sinister than the ordinary error rate of DNA replication. This is the insight that made modern tuberculosis treatment possible, and it was understood with remarkable clarity in the earliest days of chemotherapy. When streptomycin, the first effective antituberculosis drug, was introduced in the mid-1940s and given alone, patients improved dramatically for a few weeks and then relapsed with streptomycin-resistant disease. The susceptible majority had been killed, and the pre-existing resistant minority, freed from competition and no longer suppressed, had multiplied to take its place. The drug had not created resistance; it had selected for it. The British Medical Research Council's pioneering trials of the late 1940s and early 1950s established the solution, and it is the foundation of tuberculosis therapy to this day: never treat with one drug alone. If a lesion contains perhaps one organism in a hundred million resistant to drug A, and independently one in a million resistant to drug B, then the chance that any single organism is resistant to both A and B simultaneously is the product of those two small numbers — vanishingly small, smaller than the entire bacterial population. Two effective drugs cover each other's escape mutants. The resistant mutant to drug A is killed by drug B, and vice versa. Standard tuberculosis treatment uses four drugs at the outset precisely to make the simultaneous emergence of resistance to all of them effectively impossible. The corollary is the mechanism by which resistance is actually created in the real world, and it is worth stating bluntly because it is the single most important thing to understand about the origin of multidrug-resistant tuberculosis: resistance is made by effective monotherapy, whether or not anyone intended monotherapy. A patient prescribed four drugs but whose strain is already resistant to two of them is, in effect, receiving two-drug therapy, then perhaps functional monotherapy as further resistance is selected. A patient who takes their pills erratically, or whose drug supply runs out intermittently, or who is given a regimen with too few effective drugs, or who absorbs the drugs poorly because of advanced disease or diarrhoea, exposes the bacterial population to conditions in which one drug after another is, at some moment, effectively acting alone. Each time that happens, the resistant subpopulation to that drug is selected and amplified. Resistance in tuberculosis is, overwhelmingly, iatrogenic in origin — a product of how the disease is treated, not of the bacterium acting on its own. This is not a comfortable fact, because it means the resistant epidemic is, at root, a monument to decades of inadequate treatment programmes. But it is also the fact that makes resistance preventable: get the regimen right, ensure the drugs are taken, and the openings close. What the mutations actually are Because M. tuberculosis acquires resistance through chromosomal mutation rather than by importing ready-made resistance genes, each resistance phenotype corresponds to changes in a particular gene or small set of genes — and because those genes are known, they can be read directly. This is the molecular basis of the entire diagnostic revolution described in the chapters to come. A test that detects rifampicin resistance is, in the end, a test that reads a short stretch of one bacterial gene and asks whether it has been altered. Rifampicin is the clearest case and the most important. The drug works by binding to the beta subunit of the bacterial RNA polymerase, the enzyme that transcribes DNA into RNA, and jamming it. The gene encoding that subunit is called rpoB. In well over ninety-five percent of rifampicin-resistant strains worldwide, the resistance-conferring mutation lies within a single short segment of rpoB only eighty-one base pairs long — the rifampicin-resistance-determining region, universally abbreviated to the RRDR. A handful of specific codons within that small window account for the majority of all rifampicin resistance on the planet. This extraordinary concentration of resistance mutations into one tiny, predictable region is a gift to diagnostics. It means that a test does not have to survey the whole genome, or even the whole gene; it has to read eighty-one base pairs. Every rapid molecular test for rifampicin resistance, from the GeneXpert system onward, is built around interrogating the RRDR. Isoniazid, the other pillar of first-line therapy, is more complicated, which foreshadows why isoniazid resistance has always been harder to detect reliably than rifampicin resistance. Isoniazid is a prodrug: it must be activated inside the bacterium by an enzyme called catalase-peroxidase, encoded by the gene katG, before it can act. High-level isoniazid resistance usually arises through mutations in katG — most commonly at a single codon, 315 — that cripple the activating enzyme so that the prodrug is never switched on. A second route to resistance, generally conferring lower-level resistance, involves mutations in the promoter region of the inhA gene, which encodes the drug's actual target in the mycolic-acid synthesis pathway; these inhA mutations are important because they also confer cross-resistance to ethionamide, a related second-line drug. Because isoniazid resistance is spread across at least two genes and several possible mutations, no single small window captures it all, and molecular tests for isoniazid resistance are correspondingly less comprehensive than those for rifampicin. The same gene-by-gene logic applies across the pharmacopoeia. Resistance to the fluoroquinolones — the moxifloxacin and levofloxacin that are now central to several resistant-TB regimens — arises mainly through mutations in gyrA and, less often, gyrB, the genes encoding DNA gyrase, the enzyme the fluoroquinolones attack. Resistance to the injectable aminoglycosides, which dominated resistant-TB treatment for decades, maps to the rrs gene encoding the 16S ribosomal RNA and, for kanamycin specifically, to the promoter of the eis gene. Resistance to pyrazinamide typically involves the pncA gene, whose product activates that prodrug; to ethambutol, the embB gene; and, crucially for the modern regimens, resistance to bedaquiline and clofazimine is linked above all to mutations in a regulatory gene variously called Rv0678 or mmpR, which controls a drug-efflux pump. The following table summarizes the principal drug-resistance associations that underpin both diagnosis and treatment. Table 1. Principal resistance genes for key antituberculosis drugs. Drug Main gene(s) What the gene encodes Diagnostic note Rifampicin rpoB (RRDR) Beta subunit of RNA polymerase >95% of resistance in an 81-bp window; basis of all rapid tests Isoniazid katG, inhA promoter Prodrug-activating enzyme; mycolic-acid synthesis target katG 315 = high-level; inhA = low-level plus ethionamide cross-resistance Fluoroquinolones gyrA, gyrB DNA gyrase gyrA mutations predominate Aminoglycosides (amikacin, kanamycin) rrs, eis promoter 16S ribosomal RNA; acetyltransferase eis promoter specific to kanamycin Pyrazinamide pncA Prodrug-activating amidase Many scattered mutations; hard to capture Ethambutol embB Arabinosyltransferase Codon 306 common Bedaquiline, clofazimine Rv0678 (mmpR) Efflux-pump regulator Cross-resistance; central modern concern Linezolid rplC, rrl 23S ribosomal RNA and L3 protein Rare but emerging The table carries a lesson beyond its individual rows. Resistance to the older, well-characterized drugs is genetically tidy, concentrated in known hotspots that molecular tests can read with confidence. Resistance to some of the newer and the harder-to-test drugs — pyrazinamide, with its mutations scattered across the whole of pncA, or bedaquiline, where a single efflux mechanism quietly confers cross-resistance to a second drug — is messier, and that messiness is exactly where the diagnostic and stewardship challenges of the coming years will concentrate. Acquired resistance and transmitted resistance There is a distinction that runs through all of tuberculosis epidemiology and that is essential to grasp before we go further, because it determines what kind of problem resistance actually is in a given place. Resistance can be acquired, meaning it arose within an individual patient during the course of inadequate treatment, through the selection process described above. Or it can be transmitted, also called primary resistance, meaning the patient was infected from the start with an already-resistant strain breathed in from someone else. The patient with transmitted resistance never had drug-susceptible disease; they caught the resistant form directly, exactly as one catches ordinary tuberculosis. For many years the implicit mental model of multidrug-resistant tuberculosis was the acquired one: the difficult patient, the interrupted treatment, resistance generated at the bedside. That model justified a certain moralism about adherence and a focus on supervising the swallowing of pills. But molecular epidemiology, which can fingerprint strains and show when the same resistant clone is spreading through a community, has made clear that in the high-burden settings where most resistant tuberculosis occurs, the majority of multidrug-resistant disease is now transmitted, not acquired. People are not generating resistance in their own bodies so much as catching fully formed resistant strains from others, in households, on crowded transport, in prisons, in mines, in hospital waiting rooms. The resistant epidemic has become self-sustaining: it propagates itself through transmission of established resistant lineages, some of which — particular strains of the so-called Beijing lineage, for example — have spread across whole regions. This shift has profound implications, and it reframes the purpose of the rapid diagnostics that are this book's recurring theme. If most resistance were acquired, the priority would be flawless treatment of susceptible disease to stop resistance from being generated. That remains essential. But if most resistance is transmitted, then the priority becomes finding and treating the people who already have resistant disease, fast, before they infect others — because every month that a person with undiagnosed resistant tuberculosis spends coughing in a crowded home or clinic is a month of onward transmission of a strain that will be just as hard to treat in its next host. The case for diagnosing resistance in two hours rather than two months is not merely a case about getting an individual patient onto the right drugs sooner, though it is that. It is a case about interrupting the chain of transmission of resistant strains through the community. Speed of diagnosis is, in a transmitted epidemic, a form of prevention. The ladder of resistance Finally, a word about terminology, because the vocabulary of resistant tuberculosis is both important and a source of genuine confusion, having been revised more than once as the drugs and the understanding changed. The foundational category is multidrug-resistant tuberculosis, MDR-TB, defined as disease resistant to at least isoniazid and rifampicin — the two drugs that make ordinary tuberculosis curable in six months. Lose both and the whole first-line regimen collapses. In practice, because rifampicin resistance is both the most consequential loss and the one that rapid tests detect directly, programmes increasingly speak of rifampicin-resistant tuberculosis, RR-TB, as the entry point — and the two are usually managed together under the combined heading MDR/RR-TB, which is the category in which the WHO reports its global figures. The great majority of rifampicin-resistant strains turn out to be resistant to isoniazid as well, so in practical terms a rifampicin-resistance result is treated as the trigger for a resistant-TB regimen. Above MDR-TB sits a more ominous tier. Historically the term was extensively drug-resistant tuberculosis, XDR-TB, coined in 2006 to describe strains resistant not only to isoniazid and rifampicin but also to the fluoroquinolones and the injectable second-line drugs — the agents that were, at the time, the backbone of MDR-TB treatment. XDR-TB announced itself to the world that same year with a notorious outbreak at Tugela Ferry in KwaZulu-Natal, South Africa, where a cluster of HIV-positive patients fell ill with strains resistant to everything then available and died with terrifying speed, most within weeks of diagnosis. As the drug landscape changed, the definitions were revised. In 2021 the WHO re-cast the ladder around the new agents: pre-XDR-TB now denotes MDR/RR-TB with additional resistance to any fluoroquinolone, and the redefined XDR-TB denotes MDR/RR-TB with fluoroquinolone resistance plus resistance to at least one of the other key drugs that now anchor treatment, such as bedaquiline or linezolid. The shifting definitions can be dizzying, but the underlying idea is stable and simple: the more of the effective drugs a strain has defeated, the fewer options remain, the longer and harder the treatment, and the worse the outcome. In 2024 the WHO estimated that about eighteen percent of multidrug-resistant strains had already crossed into the pre-XDR tier — nearly one in five — a reminder that the ladder is not a theoretical abstraction but a population that exists, in the hundreds of thousands, right now. That ladder is the terrain on which the rest of this book operates. Every advance described in the coming chapters — the molecular tests, the toxic old injectables, the new oral drugs, the short regimens — is in some sense an answer to a question posed by the mutations and the selection pressures laid out here. The bacillus writes its resistance into a handful of genes. Medicine's recent task has been to learn to read that writing quickly, and to build regimens the bacillus cannot easily write its way out of. The Old World: Slow Diagnosis and the Injectable Era To appreciate how much has changed, you have to spend some time in the world that preceded the change, because that world was not a distant historical curiosity. It was the standard of care across most of the planet well into the 2010s, and in some places its shadow has not fully lifted. For a generation of patients and the clinicians who treated them, drug-resistant tuberculosis meant a particular and grim sequence: a diagnosis that arrived late, if it arrived at all, followed by two years of treatment with a regimen built around daily injections that could leave a person permanently deaf. Understanding that sequence in concrete detail is the only way to measure the achievements of the molecular and pharmacological revolutions that followed, and to understand why those revolutions were pursued with such urgency. Diagnosis by microscope and the long wait for culture For most of the twentieth century, and for most of the world's tuberculosis patients until very recently, diagnosis rested on a technique that Robert Koch's contemporaries would have recognized: the sputum smear. A patient coughs up sputum; a laboratory technician smears it on a glass slide, stains it by the Ziehl-Neelsen method or a fluorescent variant, and examines it under a microscope for the red, rod-shaped acid-fast bacilli that signal tuberculosis. Smear microscopy is cheap, fast, and requires only a microscope and a trained eye. It has one overwhelming virtue — it identifies the most infectious patients, those whose sputum teems with bacilli — and several crippling limitations that defined the diagnostic era. The first limitation is sensitivity. Smear microscopy can only see bacilli when they are present in large numbers — on the order of five to ten thousand organisms per millilitre of sputum. Below that threshold the smear is negative even though the patient has tuberculosis. This means smear microscopy misses a large fraction of cases: patients early in their disease, patients with HIV whose tuberculosis often produces few bacilli, and children, who rarely produce classically positive sputum at all. Across settings, smear microscopy detects only about half to two-thirds of culture-confirmed pulmonary tuberculosis, and far less of the tuberculosis in people living with HIV — precisely the population in which the disease was most lethal. The second limitation is more fundamental to this book's subject: smear microscopy tells you nothing whatsoever about drug resistance. A bacillus resistant to rifampicin looks exactly like a susceptible one under the microscope. The smear can tell you that a patient has tuberculosis; it cannot tell you which tuberculosis, or which drugs will work. For that, the old world had only one answer, and that answer was culture. To know whether a strain was resistant, you had to grow it. The classical method, phenotypic drug-susceptibility testing, means culturing the organism and then growing it again in the presence of each drug to see whether the drug stops it. Because M. tuberculosis grows so slowly, this is a process measured in weeks stacked upon weeks. Growing the organism from sputum on solid Löwenstein-Jensen medium could take three to six weeks; liquid culture systems, faster, still needed one to three. Then the drug-susceptibility test itself required another two to four weeks of growth in the presence of the drugs. From the day a patient produced sputum to the day a clinician held a result saying "resistant to rifampicin and isoniazid," two to three months might easily pass. In many high-burden settings, culture and drug-susceptibility testing were available only in a distant reference laboratory, if at all, so specimens had to be transported, results returned, patients traced. Three months was optimistic. Often the result never came, or came after the patient had been lost, or died. Consider what that delay meant in practice. A patient with undiagnosed multidrug-resistant tuberculosis, presenting with the classic symptoms, would typically be started on the standard first-line regimen — the four drugs for drug-susceptible disease — because that is what the diagnosis of "tuberculosis" called for and nothing yet said otherwise. For two or three months that patient would take a regimen to which their strain was resistant. They would not improve, or would improve and relapse. They would remain infectious, coughing resistant bacilli into their household and community. The selection pressure of a failing regimen would, in some, amplify the resistance further. And only after months of this would the culture result arrive to reveal that the patient had been receiving, in effect, no effective treatment at all. The diagnostic delay was not a neutral inconvenience. It was itself a driver of transmission, of resistance amplification, and of death. The entire case for rapid molecular diagnosis, which the next chapter takes up, is a case against this delay. The standardized regimen and its burdens Suppose, despite everything, the diagnosis of multidrug-resistant tuberculosis was made. What then? The patient entered a treatment regimen that, by any reasonable standard, was an ordeal — and that was the best medicine had to offer. The conventional regimen for MDR-TB, as codified in WHO guidance through the 2000s and into the 2010s, lasted eighteen to twenty-four months and was assembled from a hierarchy of second-line drugs, so called because they were reserved for use when the first-line drugs failed. The reason they were second-line is that they are, as a class, less effective, more toxic, and harder to take than isoniazid and rifampicin. The regimen was conventionally divided into an "intensive phase" of perhaps six to eight months, during which an injectable drug was given, followed by a "continuation phase" of a year or more on oral drugs alone. A typical regimen might combine an injectable aminoglycoside or the polypeptide capreomycin; a fluoroquinolone such as moxifloxacin or levofloxacin; and a supporting cast of older oral agents — ethionamide or prothionamide, cycloserine or terizidone, para-aminosalicylic acid, pyrazinamide, ethambutol, clofazimine — chosen to assemble at least four or five drugs likely to be effective. Read that list again and notice its character. These are not elegant modern antibiotics. Several date from the 1950s and 1960s. Para-aminosalicylic acid, a drug introduced in the 1940s, had to be taken in enormous daily doses — many grams — and routinely caused nausea and vomiting so severe that patients dreaded it. Ethionamide likewise provoked relentless gastrointestinal distress and could disturb thyroid function and mood. Cycloserine was notorious for neuropsychiatric effects: depression, psychosis, suicidality in a minority, so that patients on it needed monitoring not only of their lungs but of their minds. The pill burden was punishing — a patient might swallow fifteen or twenty tablets a day, many of them large and unpalatable, every day for the better part of two years, while also attending daily or near-daily for an injection. Simply taking the regimen was a full-time occupation and an assault on the body, undertaken by people who were often already poor, malnourished, and sick. And the results, after all that, were poor. Across the world, the treatment-success rate for multidrug-resistant tuberculosis under these regimens hovered around fifty percent, and in some programmes well below it. Half of the patients who began this two-year ordeal did not complete it successfully: they died, or the treatment failed and the disease persisted, or they were lost to follow-up — unable to sustain two years of toxic daily therapy, dropping out to return to work or family or simply to escape the misery, often with partially treated disease that risked becoming still more resistant. A cure rate of one in two, for a treatment this long and this harmful, is a figure that should be sat with. It describes a therapy that was failing half the people it was given to, and harming many of the other half along the way. That was the standard of care. Against that baseline, the modern figures — treatment-success rates approaching ninety percent in trial populations receiving six-month oral regimens — represent not an incremental improvement but a reversal of fortune. The particular cruelty of the injectables Among all the burdens of the old regimen, one stands out for its combination of severity, permanence, and the fact that it struck the young and the otherwise-recovering: the toxicity of the injectable drugs. These agents are important enough, and their removal consequential enough, that the next-but-one chapter is devoted to them in full. But their place in the old world must be established here, because the injectables were, in many ways, the defining feature of pre-modern MDR-TB treatment and the single thing patients feared most. The injectable drugs — the aminoglycosides kanamycin, amikacin, and streptomycin, and the related polypeptide capreomycin — were given by intramuscular injection, or sometimes intravenously, daily or several times a week, for months. Anyone who has received a daily intramuscular injection for even a week can imagine what months of them does to the body and the spirit. But the injections' real cruelty was not the needle. It was that these drugs are ototoxic: they poison the hair cells of the inner ear, the cochlea, that transduce sound into nerve signals. The damage is cumulative, dose-related, and — this is the essential point — permanent. A patient could begin treatment hearing normally and finish it, cured of tuberculosis, deaf. The hearing does not come back. In cohort after cohort, a substantial minority of patients on injectable-containing regimens — in some series approaching half when tested carefully with audiometry — suffered measurable hearing loss, and a meaningful fraction lost enough hearing to be disabled by it: unable to hear a conversation, cut off, in the cases of children, from language acquisition and schooling. There is a terrible irony in this. The injectables were kept in the regimen because they worked — they contributed to the bacterial kill. A patient would endure the injections, be cured of a lethal disease, and emerge with a permanent, life-altering disability inflicted not by the disease but by the treatment. The clinicians who prescribed these regimens knew this and prescribed them anyway, because the alternative was to withhold an effective drug from a patient with a frequently fatal illness. It was a defensible choice in the absence of anything better. But it was a choice made under duress, and the sense of that duress — the knowledge that the best available care routinely maimed the people it saved — hung over the field for years. When the new oral drugs made it possible to build effective regimens without injectables, the speed with which the global community abandoned the injectables reflected not only the evidence but something closer to relief. The old world, then, can be summarized in a single sentence: a diagnosis that came months too late, if at all, followed by two years of toxic treatment that cured only half and could leave the cured deaf. Every element of that sentence has now been addressed by a specific innovation. The months-long diagnosis was answered by molecular testing. The two-year duration was compressed to six months. The toxic old drugs were largely replaced. The injectables were removed outright. The chapters that follow take up each of these in turn. But none of them will register as the achievement it is unless the starting point is held clearly in mind: this is where the field was, within living clinical memory, and for a great many patients the transition out of it is still unfinished. GeneXpert and the Molecular Revolution In December 2010, the World Health Organization took the unusual step of endorsing a new diagnostic test within months of the pivotal evidence being published, and issued a formal recommendation that national programmes adopt it. The test was called Xpert MTB/RIF, and it ran on a platform, GeneXpert, made by the California company Cepheid. What made the endorsement so striking was not only its speed but what the test promised to do: to detect Mycobacterium tuberculosis and simultaneously determine whether the strain was resistant to rifampicin, directly from a sample of the patient's sputum, in under two hours, with almost no hands-on laboratory work. Set against the world described in the previous chapter — smear microscopy that saw nothing of resistance, culture that took months — this was not an improvement in degree. It was a different category of thing. The molecular diagnosis of rifampicin resistance is the hinge on which the modern management of resistant tuberculosis turns, and GeneXpert is the technology that made it real at scale. What the machine does The principle behind Xpert MTB/RIF follows directly from the biology set out two chapters ago. Rifampicin resistance is overwhelmingly concentrated in one eighty-one-base-pair window of the rpoB gene. If you can read that window and detect whether its sequence has been altered, you can detect rifampicin resistance. The test does exactly this, using a technique called real-time polymerase chain reaction — PCR — which amplifies a specific stretch of DNA from a tiny starting quantity and reports, as it goes, whether the target sequence is present. The engineering is what made it usable outside a specialized molecular laboratory. The entire process is contained within a single sealed plastic cartridge. A technician liquefies and inactivates the sputum sample with a supplied reagent, pipettes it into the cartridge, loads the cartridge into the GeneXpert machine, and presses start. Everything after that is automatic. Inside the cartridge, the sample is washed and filtered, the bacteria are broken open by ultrasonic lysis to release their DNA, and the DNA is subjected to PCR targeting the rpoB RRDR. The clever part is the detection chemistry: the cartridge contains a set of five molecular probes — fluorescent "beacons" — each designed to bind to a different overlapping segment spanning the resistance-determining region of a normal, susceptible rpoB gene. If the region is intact, as it is in a drug-susceptible strain, all five probes bind and fluoresce on schedule. If the region carries a resistance mutation, one of the probes fails to bind properly over its altered segment, and its fluorescence is delayed or absent. The machine reads this pattern and reports one of a few plain answers: M. tuberculosis detected or not; and if detected, rifampicin resistance detected, not detected, or indeterminate. No gel, no sequencing, no interpretation of raw data by the operator. A result in roughly ninety minutes to two hours. The implications for the delay problem were immediate and profound. A patient who, in the old world, would have been started on first-line drugs and left for three months before a culture revealed resistance could now, in principle, be identified as rifampicin-resistant on the day they first produced sputum, and started on an appropriate regimen within days. The months of ineffective treatment, of continued transmission, of resistance amplification, of clinical deterioration — the entire cost of the diagnostic delay — could be collapsed to almost nothing. And because the test detects tuberculosis itself with far greater sensitivity than smear microscopy, it also found cases the microscope missed, including in people with HIV, the population in whom the old smear had been most blind. From a tool to a programme A diagnostic test is only as good as its reach, and the history of GeneXpert is as much a story of deployment as of invention. The initial WHO endorsement was followed by an enormous, donor-supported effort to place GeneXpert machines in laboratories across the high-burden world. Negotiated price reductions brought the cost of a cartridge down from its original level to a concessional price for eligible countries, though even the reduced price remained a real constraint on the volume of testing programmes could afford. Over the following years, tens of thousands of GeneXpert modules were installed across Africa, Asia, Eastern Europe, and Latin America, and the test moved from reference laboratories toward district and even peripheral health facilities. The machine's relative robustness — it needs a stable power supply and a controlled temperature, but not the full infrastructure of a molecular laboratory — made this decentralization possible in a way that culture never could have been. The WHO's position hardened over the decade from recommending the test as an option to recommending it as the initial diagnostic test for all people with signs and symptoms of tuberculosis, displacing smear microscopy from its century-long role as the front door to diagnosis. This was a genuine inversion of the diagnostic logic. For a hundred years, the first question had been "does this person have tuberculosis?", answered by the microscope, with the question of resistance deferred to a slow and often unavailable second step. Now the two questions were answered at once, at the point of first contact: tuberculosis, yes or no, and rifampicin resistance, yes or no, together. Resistance detection ceased to be a specialized afterthought and became a routine part of diagnosis. That reconception — testing everyone for resistance up front, rather than waiting for treatment to fail — is perhaps the single most important programmatic consequence of the molecular era, and it is why rifampicin resistance, detectable by Xpert, became the practical entry point to the whole resistant-TB treatment pathway. Yet the reach remains incomplete, and the incompleteness is one of the central unfinished stories of the field. Despite fifteen years of deployment, the WHO reported that in 2024 only about fifty-four percent of people diagnosed with tuberculosis worldwide had been tested with a WHO-recommended rapid molecular test as their initial diagnostic. In other words, nearly half of the world's tuberculosis was still being diagnosed by methods that cannot see resistance at all. Every one of those cases is a patient whose resistance, if present, will again be discovered only when treatment fails — the old delay, reproduced in the present. The technology to abolish the diagnostic delay has existed since 2010. The delay persists because the technology has not reached everyone, and the reasons it has not — cartridge cost, machine maintenance, supply chains, the recurring expense that every test represents to a strained budget — are precisely the systemic constraints that the closing chapters of this book argue are now the binding limit on progress. Ultra, Truenat, and the widening molecular toolkit Xpert MTB/RIF was the beginning, not the end, of molecular tuberculosis diagnostics, and the toolkit has broadened in ways that matter for resistant disease. In 2017 Cepheid released a next-generation cartridge, Xpert MTB/RIF Ultra, which the WHO reviewed and recommended as at least equivalent and in important respects superior. Ultra was designed to address the main weakness of the original: its sensitivity in the hardest cases — paucibacillary disease, where few organisms are present, as in people with HIV, in children, and in tuberculosis outside the lungs. Ultra achieves a markedly lower limit of detection, able to find far fewer bacilli per millilitre than the original cartridge, by adding two multicopy amplification targets and by using a different method — melting-temperature analysis — to read the rpoB region for resistance. The improved sensitivity came with a characteristic trade-off: Ultra introduced a new semiquantitative category, "trace," for samples with the very lowest bacterial loads, which is highly sensitive but can be difficult to interpret, especially in people with a history of previously treated tuberculosis in whom dead bacillary DNA may linger. Alongside the Cepheid platforms, other molecular technologies have been developed and endorsed, loosening the dependence on a single manufacturer that had long made the field nervous. Truenat, a chip-based real-time PCR system developed by the Indian company Molbio, performs a comparable function — detecting tuberculosis and, in a follow-on assay, rifampicin resistance — on a small, battery-capable, portable device suited to peripheral settings with limited infrastructure. The WHO's endorsement of Truenat gave programmes a second validated option and introduced a measure of competition into a market that had been effectively a monopoly. For the detection of resistance beyond rifampicin, line-probe assays — strip-based molecular tests, such as the GenoType MTBDR line from the German company Hain — remain in use; these can detect resistance to isoniazid, to the fluoroquinolones, and to the second-line injectables by reading the relevant genes, though they generally require a culture or at least a smear-positive specimen and more laboratory handling than the cartridge-based tests. And at the frontier, targeted next-generation sequencing assays, such as the Deeplex Myc-TB assay endorsed by the WHO in 2023, can read many resistance-associated genes at once directly from sputum, producing something approaching a comprehensive resistance profile in a single run — a capability that will become increasingly important as regimens come to depend on knowing a strain's susceptibility to bedaquiline, linezolid, and the fluoroquinolones, not merely to rifampicin. The following table sets the principal diagnostic approaches side by side. Table 2. Principal tuberculosis diagnostic methods and their resistance-detection capabilities. Method Time to result Detects rifampicin resistance Detects other resistance Typical setting Sputum-smear microscopy Same day No No Any laboratory Solid/liquid culture 1–6 weeks Via follow-on phenotypic DST (weeks more) Yes, phenotypic Reference laboratory Xpert MTB/RIF ~2 hours Yes (rpoB RRDR) No District to peripheral Xpert MTB/RIF Ultra ~2 hours Yes, higher sensitivity No District to peripheral Truenat MTB / MTB-RIF Dx ~1 hour Yes (follow-on assay) No Peripheral, portable Line-probe assay (LPA) ~1–2 days Yes Isoniazid, fluoroquinolones, injectables Reference laboratory Targeted NGS (e.g. Deeplex) ~1–2 days Yes Broad multi-drug profile Reference / sequencing lab As the table shows, no single test yet does everything. The cartridge-based assays are fast and decentralizable but see only rifampicin; the sequencing-based assays see almost everything but need more infrastructure and time. The direction of travel is clear — toward rapid, decentralized tests that report a fuller resistance profile — but the present reality is a layered system in which a rapid rifampicin result at the front line triggers referral for more comprehensive testing of the drugs that the modern regimens depend on. The limits of reading rifampicin For all its transformative power, the molecular detection of rifampicin resistance carries pitfalls that a careful clinician must understand, because taking a GeneXpert result as infallible can harm patients in both directions. The test reads a genotype and infers a phenotype, and the inference is usually but not always correct. The most important false-positive problem concerns what are called silent, or synonymous, mutations. The probes in the original Xpert cartridge detect any change in the rpoB region that disrupts probe binding — but not every change in that DNA sequence actually confers resistance. Some mutations alter the DNA without altering the resulting protein in a way that matters for drug binding, or alter it in a way that the probe reads as abnormal while the strain remains clinically susceptible to rifampicin. When the test reports resistance on the basis of such a change, a genuinely susceptible patient may be diverted onto a long, toxic, expensive resistant-TB regimen they never needed. The magnitude of this problem depends on the local prevalence of resistance — in a setting where true rifampicin resistance is rare, a larger proportion of positive results will be false — which is why confirmatory testing, by sequencing or phenotypic methods, is recommended particularly where resistance is uncommon and the pre-test probability is low. The converse error also exists. Because the test reads only the eighty-one-base-pair window, it will miss the small minority of rifampicin-resistant strains whose resistance mutation lies outside the RRDR — rare, but real. It can also be confounded by mixed infections, in which susceptible and resistant populations coexist, and by the "hetero-resistance" that arises during the emergence of resistance, where the resistant subpopulation is too small for the test to call confidently. And a rifampicin result, however accurate, says nothing about isoniazid, the fluoroquinolones, bedaquiline, or linezolid — the drugs that determine whether a modern regimen will actually work. A patient can be correctly identified as rifampicin-resistant and still be carrying a strain with additional resistance that the front-line test cannot see, which is why the molecular detection of rifampicin resistance is the beginning of the diagnostic workup for resistant disease, not the end of it. None of these caveats diminishes the achievement. They refine it. The molecular revolution did not abolish the need for judgement, for confirmatory testing, for a laboratory network capable of profiling resistance to the drugs that now matter most. What it abolished was the three-month void in which a resistant patient received no effective treatment and infected everyone around them. That void was the central scandal of the old world. GeneXpert and the technologies that followed it closed it — wherever they reach. The task that remains, and it is a formidable one, is to make them reach everyone. Hashtags: #Tuberculosis #MultidrugResistantTuberculosis #MDRTB #RifampicinResistantTuberculosis #DrugResistantTuberculosis #RapidMolecularDetection #GeneXpert #XpertMTBRIF #XpertUltra #Truenat #MolecularDiagnostics #RifampicinResistance #rpoB #DrugSusceptibilityTesting #BPaLM #Bedaquiline #Pretomanid #Linezolid #Moxifloxacin #ShortCourseTreatment #AllOralRegimens #InjectableFreeTreatment #AntimicrobialResistance #TuberculosisStewardship #FutureOfTuberculosisTreatment

Download the Book (PDF): Introduction Every tunnel begins as an act of disturbance. Before the first cut, the ground around the future opening sits in a balance it has held for thousands or millions of years: grains pressed against grains, water filling the spaces between them at whatever pressure the water table above dictates, the weight of the overburden carried evenly down through the soil column. The moment a cutter or an excavator removes material, that balance is gone. The ground around the hole has lost the support it was leaning on, and it begins, immediately and without negotiation, to move toward the void. Most people imagine tunneling as a contest of strength: a machine that is powerful enough to grind through rock, a lining that is thick enough to hold up the city above. Strength matters, but it is not what decides whether a tunnel in soft ground succeeds or fails. What decides it is control. The successful tunnel engineer controls two quantities, pressure and time, and refuses to let the ground take charge of either. The failures that fill the history of the industry, from crater collapses in city streets to flooded headings and abandoned machines, are almost always failures of that control rather than failures of material strength. The ground was not too strong for the tools. It was given the wrong pressure, or too much time, or water was allowed to change the rules. That is the argument of this booklet. Soft-ground tunneling is the management of a temporary equilibrium between the ground, the groundwater and the support, and the two dominant modern methods are best understood as two different strategies for managing it. The Earth Pressure Balance tunnel boring machine controls pressure: it keeps the face supported continuously, at a pressure matched to the ground and water in front of it, so that the ground never gets the chance to move. The New Austrian Tunneling Method, with its sprayed concrete linings, controls time and deformation: it lets the ground move a little, deliberately, so that the ground carries part of its own load, then stiffens the support at the right moment before the movement becomes harmful. Groundwater is the variable that sits across both strategies and most often breaks them, because water under pressure turns ground that would stand into ground that flows. Why soft ground is the hard case Rock tunneling has its own dangers, including rockbursts, squeezing ground, fault zones and the occasional catastrophic inflow, but intact rock usually has enough strength to stand unsupported for some period after excavation. A great deal of hard-rock tunneling consists of cutting the hole and then installing support at a steady, deliberate pace. Soft ground offers no such grace. Sands, silts, clays, gravels and the mixed glacial and alluvial deposits beneath most of the world's large cities have little or no strength of their own when confining pressure is removed. Dry sand runs. Saturated sand flows. Soft clay squeezes into the opening. Stiff clay stands for a while and then fails along fissures that were invisible on the day of excavation. Soft ground is also where most urban tunnels are built, because cities grow on river plains, estuaries and coastal deposits. The metro lines of London, Paris, Shanghai, Singapore, Los Angeles and Mexico City, the sewer and storm relief tunnels beneath almost every large city, and the road tunnels threaded under waterfronts all pass through soils rather than rock for much of their length. That places the most demanding ground conditions directly beneath the most sensitive structures: buildings on shallow foundations, Victorian brick sewers, gas mains, historic churches, live railways. The tolerance for ground movement is measured in millimeters, and the consequence of losing control is measured in collapsed streets and, occasionally, in lives. The two strategies The closed-face tunnel boring machine is the dominant answer to that problem for long, continuous drives. A modern Earth Pressure Balance (EPB) machine is a steel cylinder, typically between five and fifteen meters in diameter, with a rotating cutterhead at the front and a sealed chamber behind it. The chamber fills with excavated soil, which is kept under pressure by controlling the rate at which a screw conveyor removes it. The pressurized soil pushes against the face, balancing the earth and water pressure in front of the machine. Behind the cutterhead, the machine erects rings of precast concrete segments and pushes itself forward off them with hydraulic jacks, while grout is injected into the gap between the rings and the excavated ground. Done well, the process is continuous: the face is never left unsupported, the ground is never given time to move, and the surface above may settle by only a few millimeters. The New Austrian Tunneling Method, known in Britain and elsewhere as the sprayed concrete lining (SCL) method, takes an older and more flexible approach. The ground is excavated in sequences, often by excavator rather than by machine, and the newly exposed surface is sprayed with concrete within hours or minutes. The lining starts soft and stiffens as it cures, while instruments track how much the opening is converging. The method relies on the ground's ability to redistribute stress around the opening, forming a load-bearing ring of ground that works together with the thin concrete shell. It is the method of choice where tunnels change shape, where they are too short to justify a machine, and where openings must be cut into the side of a bored tunnel: station caverns, junctions, cross passages, crossovers, shafts and adits. Its advantages are adaptability and economy. Its weakness is that it depends on judgment at the face, and when that judgment fails, the failure can be sudden. Most large urban projects now use both. A metro line is bored with EPB machines between stations, and the stations, cross passages and junctions are built by sprayed concrete methods, often from inside the bored tunnels. Understanding either method in isolation misses how the two are combined, and how the interfaces between them, where a concrete segment ring is cut open so that a hand-excavated passage can begin, are among the most hazardous operations in the whole project. Water as the deciding variable Groundwater runs through both strategies. In a soil below the water table, the water in the pores carries part of the load. When a tunnel lowers the pressure at the face, water flows toward the opening, and that flow exerts forces on the soil grains that can carry them out with it. Sand that would stand at a steep angle when dry or damp can run like liquid when water is moving through it. Fine silts, which drain slowly, can turn to slurry. Under a river or a sea, where the hydrostatic head may be tens of meters, the water pressure alone can exceed the pressure that any open-face method can resist. For an EPB machine, water is a problem of sealing and of muck control: the screw conveyor must hold back water under pressure while discharging excavated soil, and if the soil in the chamber is too permeable, water and fines will blow through the screw. For sprayed concrete tunneling, water is a problem of stability and of lining design: the ground must be drained, grouted or frozen before it is opened, and the finished lining must either keep water out entirely or be designed to collect and carry it away. For the people who must enter the excavation chamber of a machine to change worn cutting tools, water pressure is a direct physiological hazard, because the chamber must be held at a pressure high enough to keep the ground stable, and human beings must breathe and decompress under it. What follows The chapters run in the order an engineer encounters the problems. The first explains how ground behaves when a hole is cut into it, and why the presence of water changes that behavior so radically. The second explains how a method is chosen, because the most consequential decision on any soft-ground project is taken before construction begins. Three chapters then follow the EPB machine in detail: how it holds the face, how the excavated soil is conditioned so that it can hold pressure at all, and how the machine is driven so that settlement stays within limits and the segmental lining is built correctly. Two chapters treat the New Austrian Tunneling Method, first as a principle of ground-support interaction and excavation sequencing, then as a material practice centered on sprayed concrete and its early strength. The eighth chapter takes on groundwater ingress under high hydrostatic pressure directly: seepage, inflow, dewatering, grouting, freezing and waterproofing. The ninth follows the people and operations at the margins of the closed face, including hyperbaric interventions, break-ins and cross passages, where the two methods meet. The tenth draws these strands together into the discipline that binds them, risk management and the observational method, and the conclusion asks what all of this implies for how tunnels should be specified, built and judged. No prior engineering knowledge is assumed. Terms are defined as they arise, and the few equations that matter are explained in words. The reader who finishes should be able to look at a tunnel project, whether a new metro line, a sewer under a river or a highway beneath a harbor, and understand where its real risks lie, why its builders chose the methods they did, and what they must watch every hour of every shift to keep the ground where it belongs. Chapter 1. What the Ground Does When You Open It A tunnel engineer works with two materials. One is manufactured: steel, concrete, grout and the machinery that places them. Its properties are specified, tested and certified. The other is the ground, which nobody designed, which varies over a few meters, and which is known only through a scatter of boreholes that together sample perhaps a millionth of the volume the tunnel will pass through. Every method in this booklet is a way of making the manufactured material and the natural one work together. Before looking at any method, it helps to understand what the natural material does when a hole is cut into it, because every decision follows from that response. Stress, arching and the loss of confinement Imagine a point in the ground twenty meters below a city street. The soil at that point carries the weight of everything above it: about twenty meters of soil, each cubic meter weighing roughly eighteen to twenty-one kilonewtons, plus whatever buildings stand on the surface. The vertical stress is therefore of the order of four hundred kilopascals, about four times atmospheric pressure. The horizontal stress is usually smaller in recently deposited soils, perhaps half the vertical stress, but in heavily overconsolidated clays such as London Clay, which were once buried under much greater thicknesses of sediment that have since eroded away, the horizontal stress can be equal to or greater than the vertical. The ratio of horizontal to vertical stress, called the coefficient of earth pressure at rest and written K0, is one of the first numbers a designer wants to know, because it governs both the loads on the lining and the way the ground deforms. When a circular opening is excavated at that point, the stresses that used to pass straight through the space now occupied by the tunnel must go around it. The ground immediately next to the opening loses its radial confinement entirely, since there is nothing inside the hole to push back, while the tangential stress around the opening rises, typically to about twice the original value in an elastic material. This redistribution is called arching. In a strong material, arching is benign: the ring of ground around the tunnel carries the increased tangential stress and the opening stands. In a weak material, the increased stress exceeds the strength of the ground near the opening, a zone of yielding forms, and the ground moves inward until either the support takes the load or the yielded zone grows so large that the opening collapses. Two points follow from this picture, and they run through every later chapter. The first is that the ground is not simply a load to be held up. It is a structure in its own right, capable of carrying most of the redistributed stress if it is not allowed to degrade. A soft-ground tunnel lining typically carries far less than the full weight of the overburden, because arching in the surrounding ground carries the rest. The second point is that this arching capacity is not guaranteed. It depends on the ground keeping its strength, which depends on how much it is allowed to deform, how long it is left unsupported, and whether water is flowing through it. A practical vocabulary of ground behavior Soil mechanics provides the formal theory, but tunnelers have long needed a more direct language for what ground does at an open face in the minutes and hours after excavation. The most durable vocabulary came from Karl Terzaghi, who classified ground by its behavior in the heading in work published in 1950, drawing on observations made during construction of the Chicago subway in the 1940s. The terms remain in use because they describe what a miner actually sees, and they form the basis of the classification later refined by Robert Heuer for tunnel design in North America. Table 1 sets out the main categories. Table 1. Terzaghi's tunnelman's ground classification and typical soils. Behavior What happens at an open face Typical soils Firm Face and walls stand unsupported long enough to excavate a round and install support Stiff clays above the water table, cemented sands Raveling Flakes or chunks fall from the roof and face after a delay; slow or fast raveling depending on rate Stiff fissured clays, slightly cohesive sands Running Dry granular material flows at its angle of repose until it chokes itself; "cohesive running" stands briefly first Clean dry sands and gravels, damp sands with slight apparent cohesion Flowing Saturated material moves into the opening as a viscous fluid, invading from floor, face and crown Silts, fine sands and gravels below the water table Squeezing Ground advances slowly into the opening without visible fracturing, through plastic deformation Soft to medium clays, depending on depth and stress Swelling Ground expands into the opening by absorbing water Clays with swelling minerals, such as montmorillonite Source: after Terzaghi (1950) and Heuer (1974). Two features of this table deserve emphasis. The same soil can occupy different rows depending on its water content and the stress it is under. A fine sand above the water table, slightly damp, may show cohesive running, standing for a minute before collapsing. The same sand below the water table, with water flowing toward the excavation, will flow. And a clay that is firm at ten meters may squeeze at forty, because the stress around the opening has risen relative to its strength. Ground behavior is therefore not a property of the soil alone. It is a property of the soil, its water and its stress state, taken together. The second feature is the role of time. Raveling, squeezing and swelling all happen over a period ranging from minutes to months. The stand-up time, meaning how long an unsupported span can stand before it starts to fail, is the variable that every tunneling method exploits or defends against. An open-face method relies on the stand-up time being longer than the time needed to install support. A closed-face machine denies the ground any unsupported time at all. Effective stress, and why water changes everything The most important idea in soil mechanics is Terzaghi's principle of effective stress, and no explanation of soft-ground tunneling makes sense without it. A saturated soil consists of a skeleton of solid grains with water filling the pores between them. When a load is applied to the soil, part of it is carried by the grains pressing against each other and part by the pressure in the pore water. Only the part carried through the grain contacts, the effective stress, gives the soil its strength. Water cannot carry shear, and friction between grains depends on how hard they are pressed together. If the pore water pressure rises while the total load stays constant, the effective stress falls and the soil weakens. If the pore water pressure rises until it equals the total stress, the effective stress reaches zero and a granular soil has no strength at all. This is the condition behind quicksand and liquefaction, and behind a flowing tunnel face. Below the water table, the pore water pressure at a given depth is roughly equal to the weight of the column of water above that point. At twenty meters below the water table it is about two hundred kilopascals, or two bar. At forty meters, beneath a river or an estuary, it is four bar. These pressures do not care whether the ground is strong or weak. They act on the support, on the face and on the lining with the same force everywhere at that depth, and they push water toward any opening where the pressure is lower. When a tunnel is opened below the water table and the face is at atmospheric pressure, a hydraulic gradient forms toward the excavation. Water flows through the soil toward the face, and that flow exerts a drag on the soil grains, called seepage force, which is directed toward the opening. Seepage force adds to the destabilizing effect of the lost confinement. In a clean sand, a modest inflow gradient is enough to reduce the effective stress at the face to zero, and the face flows. In a stratified deposit where a thin sand layer is sandwiched between clays, the sand layer may act as a pipe, delivering water and then sand into the tunnel while the clays above and below stand perfectly well. Many of the worst soft-ground incidents began exactly this way, with an unrecognized water-bearing lens rather than with the bulk of the ground at the face. Permeability, the ease with which water moves through a soil, therefore matters as much as strength. Clean gravels have permeabilities of the order of one centimeter per second or more; clean sands perhaps a thousandth of that; silts a thousandth less again; clays a million times less than sands. This range, covering ten orders of magnitude, is larger than for almost any other engineering property. It means that in a clay, water hardly moves at all during the hours of excavation, and the clay behaves in an undrained way: its pore pressure falls when it is unloaded, temporarily increasing its effective stress and stiffening it. That is why stiff clays such as London Clay can stand at an open face for hours. The temporary suction is a gift, but it is temporary. As water slowly migrates in to equalize the pore pressure, the clay softens. A London Clay face that stands for a shift may begin to fail if it is left over a weekend. In a sand, by contrast, drainage is effectively instantaneous on construction timescales. There is no temporary suction to rely on, only friction, and the friction depends on effective stress, which water flow destroys. This single contrast between clays that respond undrained and sands that respond drained explains much of the difference in how the two are tunneled. Volume loss and settlement at the surface However well a tunnel is built, the ground around it moves. The excavated volume is always slightly larger than the volume of the finished tunnel, because the ground moves inward at the face, into the gap around the machine's shield, and into the space between the shield and the lining before grout fills it. Tunnel engineers express this extra volume as a percentage of the tunnel's cross-sectional area and call it volume loss. If a tunnel six meters in diameter has a cross-section of about twenty-eight square meters and a volume loss of one percent, an extra 0.28 cubic meters of ground has moved toward the tunnel for every meter of its length. In clays, which do not change volume quickly, that lost volume propagates to the surface almost unchanged and shows up as a trough of settlement above the tunnel. In 1969 Ralph Peck observed that the cross-section of such troughs closely follows the shape of a normal probability curve, the Gaussian bell curve inverted. The width of the trough is described by the distance from the tunnel centerline to the point of inflection, conventionally written i. Work by O'Reilly and New in 1982, based on field records from British tunnels, showed that i is roughly proportional to the depth of the tunnel axis, with a constant of about 0.5 in clays and lower values, roughly 0.25 to 0.35, in sands and gravels. Sands therefore give narrower, deeper troughs for the same volume loss, while clays spread the settlement more widely. These empirical relationships remain the backbone of settlement prediction in practice. They allow a designer to estimate, before construction, how much each building along the route will settle and how much it will tilt or be stretched, and so to rank buildings by risk. A widely used framework, developed for the Jubilee Line Extension in London in the 1990s and building on work by Burland and colleagues, assesses buildings in stages: first a simple screen to exclude buildings where predicted settlement is negligible, then a calculation of the tensile strain each remaining building would suffer if it followed the ground movement, and finally detailed analysis of the most vulnerable structures. Buildings predicted to suffer strains in the range associated with moderate or worse damage receive protective measures such as compensation grouting, in which grout is injected beneath the foundations during tunneling to replace the lost volume as it occurs. The practical meaning of volume loss is this: it is the most direct measure of how well the tunneling method has controlled the ground. Modern EPB machines in suitable ground routinely achieve volume losses below one percent, and well-run drives in London clays and similar deposits have recorded values around half a percent or less. Open-face and sprayed concrete methods in good clays can achieve similar results; in poor ground, without adequate face control, they may exceed two or three percent, and an uncontrolled face collapse can produce a local volume loss many times larger, propagating to the surface as a sinkhole. The face, the shield and the tail Ground movement around a tunnel is easiest to understand by following a single point in the ground as the tunnel approaches and passes. Long before the face arrives, the ground ahead of it begins to relax, because the stress that used to pass through the future tunnel starts to redistribute. Measurements around tunnels typically show that some fraction of the eventual settlement has already occurred by the time the face passes directly beneath a surface point. How large that fraction is depends on how well the face is supported. An open face, or a closed face held at too low a pressure, lets the ground ahead move toward the excavation. A closed face held at the right pressure allows almost no movement ahead of the face. As the face passes, the ground meets the machine's shield or, in sprayed concrete tunneling, the freshly exposed surface. A shield is slightly smaller than the excavated diameter at the cutterhead, by a few centimeters, both to reduce friction and because the cutterhead's outermost tools, the gauge cutters, cut a slightly oversized hole. The shield may also taper slightly toward the rear. Ground can move into this annular gap, and in soft clays it does. Then, as the machine moves on, the lining ring it has built emerges from the back of the shield, the tail, leaving a further gap between the outside of the ring and the excavated ground. This tail void can be ten to fifteen centimeters or more, and if it is not filled immediately with grout, the ground will close it. Each of these three locations, the face, the shield and the tail, contributes to the total volume loss, and each is controlled by a different operational variable: face pressure, shield overcut and lubrication, and tail grouting respectively. In sprayed concrete tunneling, the equivalent locations are the face, the unsupported span behind it, and the period while the young lining is gaining stiffness. Chapters 3 to 7 take these up in turn. For now, the essential point is that ground movement is not one thing but a sum of contributions, each of which the builders can influence. Long-term behavior Ground movement does not stop when the tunnel is complete. In clays, the pore pressures disturbed by construction take months or years to return to equilibrium, and as they do, the clay consolidates or swells. A tunnel that leaks acts as a drain, lowering the pore pressure in the surrounding clay and causing long-term consolidation settlement that can eventually exceed the short-term settlement from construction. Records from several older tunnels in soft clays, including in Shanghai and London, show settlements continuing for decades after construction where linings allow seepage. Conversely, a watertight lining in clay will eventually carry the full hydrostatic pressure of the groundwater as pore pressures recover around it. This long-term picture explains why water management cannot be treated as a construction problem alone. The decision about whether a tunnel will be fully watertight, or designed to drain, affects the design load on the lining, the long-term settlement of the ground, the operating cost of pumping, and the durability of the structure over a design life that for major infrastructure is now commonly specified as one hundred and twenty years. What follows from the ground response The ground's response to an opening gives the engineer three things to control. The first is the support pressure at the face and around the perimeter, which determines how much the ground can move and whether it remains stable. The second is time, which determines how far time-dependent processes such as raveling, squeezing, drainage and softening can proceed before the support is in place. The third is water, whose pressure and flow can reduce the ground's strength to nothing. Each tunneling method controls these three differently. The closed-face machine acts on pressure, maintaining support at the face continuously and grouting the tail void immediately, so that neither time nor water flow has a chance to act. The sprayed concrete method acts on time and deformation, accepting a small and controlled movement so that the ground can arch, and placing support quickly enough to stop movement before the ground loses strength. Both methods depend on water being controlled, whether by pressure at the face, by drainage, by grouting or by freezing. The rest of this booklet follows those three controls through the two methods. Before any of that, however, the builders must decide which method to use, and that decision begins with understanding the ground well enough to know which rows of Table 1 the tunnel will pass through. Chapter 2. Choosing the Method Before the First Cut The most consequential decisions on a soft-ground tunnel are made years before any ground is excavated. Once a tunnel boring machine has been designed, manufactured and lowered into its launch shaft, its type, diameter and cutterhead configuration are fixed for the whole drive, which may run for many kilometers through ground that varies continuously. A machine chosen for the wrong ground cannot be swapped out halfway. It can be modified, sometimes at great cost and after long delays, or it can be abandoned. The choice between a closed-face machine and a sprayed concrete method, and between the two main families of closed-face machine, is therefore a bet placed on the ground model, and the quality of that bet depends on the quality of the site investigation behind it. Knowing the ground Site investigation for a tunnel is unusually difficult because the structure is long, deep and linear. A building sits on a footprint that can be drilled densely. A tunnel ten kilometers long passes through ground that can be sampled only at intervals, often limited by access in a city where boreholes must be drilled in streets, parks and private land. Typical practice places boreholes every fifty to two hundred meters along the alignment, offset from the tunnel line itself so that they do not create a pathway for water or pressurized material into the future tunnel. Every borehole drilled close to the alignment must be carefully backfilled with grout, because an open or poorly sealed borehole can become a conduit through which face pressure escapes to the surface or water reaches the machine. The investigation must answer several distinct questions. The first is stratigraphy: which soil units the tunnel will pass through, where their boundaries lie, and how variable they are. A tunnel horizon that passes from clay into sand, or that runs along the boundary between the two, is far more demanding than one that stays within a single uniform unit. Mixed-face conditions, in which part of the face is in one material and part in another, are among the most difficult for any machine, because the cutterhead and face support must cope with both at once. The second question is groundwater: the water table, any perched water above clay layers, artesian pressures in confined aquifers, the permeability of each unit and how the water pressure varies with depth and season. The third is the presence of obstructions: boulders in glacial deposits, old piles and foundations, abandoned wells, buried timber, unexploded ordnance in some cities, and the remains of earlier tunnels. The fourth question is the one most specific to closed-face machines, namely the properties that govern how the excavated soil will behave once it is inside the machine. These include grain size distribution, plasticity and consistency of fines, abrasiveness, the content of quartz and other hard minerals that wear cutting tools, and the tendency of clays to stick to steel. None of these matter much to a designer of foundations. All of them matter enormously to whether an EPB machine can maintain face pressure and how often its tools must be changed. Despite its importance, site investigation is routinely undersized. A recurring finding of reviews of tunnel cost overruns is that unforeseen ground conditions account for a large share of claims and delays, and that more and better investigation would have paid for itself many times over. Guidance from bodies such as the International Tunnelling and Underground Space Association, known as ITA-AITES, and national guidance such as the British Tunnelling Society's publications, consistently urges owners to invest more in investigation for underground works than for surface construction of similar value. The difficulty is that investigation is spent early, by the owner, when budgets are under pressure and the benefits are uncertain, while the costs of inadequate investigation fall later, often on the contractor first and then back on the owner through claims. Sharing the risk of the unknown ground Because no investigation can fully characterize the ground, the contract must decide who carries the risk of the ground being different from what was expected. Over the past four decades the tunneling industry has developed a distinctive answer to this question, most fully in the United States: the geotechnical baseline report. The baseline report sets out, in quantitative terms, the ground conditions that the contractor should assume when pricing and planning the work. It might state, for example, that a given length of tunnel will encounter boulders larger than three hundred millimeters at a stated frequency, or that groundwater inflows at the face will not exceed a stated rate, or that the tunnel will pass through clay with an undrained strength between stated limits. If the actual conditions are materially worse than the baseline, the contractor is entitled to additional payment; if they are within the baseline, the contractor bears the cost. The value of this approach is not that it predicts the ground correctly. It is that it converts a scientific uncertainty into a contractual allocation that both parties can price. Without a baseline, contractors must either price the worst case, raising costs on every project, or price an optimistic case and then fight over every unexpected boulder. The Joint Code of Practice for Risk Management of Tunnel Works in the UK, first published by the British Tunnelling Society and the Association of British Insurers in 2003, and the international version developed with the International Tunnelling Insurance Group in 2006, made formal risk registers and the clear allocation of ground risk conditions of insurability. Those codes emerged directly from a run of costly tunnel failures in the 1990s and early 2000s, which had led insurers to the edge of withdrawing from the market. The closed-face family When the decision is made to use a closed-face machine, the main choice is between earth pressure balance and slurry. Both support the face with pressure, and both erect a segmental lining behind the shield. They differ in what medium they use to transmit that pressure, how they remove the excavated material, and in the ground for which each is naturally suited. An EPB machine supports the face with the excavated soil itself. The cutterhead mixes the spoil in the excavation chamber, often with added conditioning agents, into a paste whose pressure is controlled by the rate at which the screw conveyor removes it and the rate at which the machine advances. The spoil leaves the screw at atmospheric pressure and is carried out of the tunnel on a conveyor belt or in muck cars. Because the face is supported by the soil itself, the method works best in soils that can be transformed into a soft, plastic, relatively impermeable paste: clays, silts and silty sands. Clean sands and gravels with little fine content are more difficult, because they are too permeable and too frictional to form a good plug in the screw. A slurry machine supports the face with a pressurized bentonite suspension. The bentonite slurry penetrates a short distance into the soil at the face and forms a thin, low-permeability layer called a filter cake, through which the slurry pressure is transmitted to the soil skeleton. The excavated soil mixes with the slurry and is pumped out of the tunnel through pipes to a separation plant at the surface, where sand and gravel are removed by screens and hydrocyclones and the slurry is recycled. Slurry machines work best in permeable granular soils, sands and gravels, where the filter cake forms well and where separation is easy. They are less suited to clays, which disperse into the slurry and are expensive to separate, and which can clog the machine's chamber and pipes. The boundary between the two types has shifted over the past three decades. Advances in soil conditioning, discussed in Chapter 4, have extended EPB machines into sandier and more permeable ground than was once thought possible. Variable density machines, developed by the German manufacturer Herrenknecht in the 2010s, can switch between EPB, slurry and an intermediate high-density slurry mode within a single drive, and have been used on projects in Kuala Lumpur and elsewhere where karstic limestone and highly variable ground posed problems for either single mode. Mixshield machines, a form of slurry machine with an air bubble controlling slurry pressure in the chamber, are widely used in high water pressure conditions under rivers and seas. Even so, the general guidance remains useful. The fine content of the soil, meaning the proportion passing a sieve of about 0.06 millimeters, is the single most useful indicator. Where fines exceed roughly 20 to 30 percent, EPB is usually the natural choice. Where fines are below about 10 percent and permeability is high, slurry is usually preferred, unless heavy conditioning can make EPB work. Between those bounds lies a zone where either can work and the choice depends on the water pressure, the availability of space for a slurry treatment plant, the cost of spoil disposal, and the contractor's experience. The sprayed concrete alternative For some tunnels, no machine makes sense. A tunnel boring machine is a large capital investment, commonly costing tens of millions of dollars for a metro-sized machine, with months of lead time for design and manufacture and additional months of launch preparation. That investment pays off over long drives of uniform cross-section. For short tunnels, for tunnels whose cross-section changes, and for the large and complex openings that every underground railway contains, a method based on sprayed concrete is usually cheaper and more practical. The sprayed concrete lining method, known internationally under the name New Austrian Tunneling Method, involves excavating the ground in small steps, by excavator, roadheader or drill and blast, and supporting each newly exposed area with sprayed concrete, often reinforced with steel mesh or fibers and sometimes supplemented by lattice girders, rock bolts or spiles. Because the lining is sprayed rather than assembled, it can take almost any shape: the wide flat-bottomed profile of a station platform tunnel, the bell-mouth junction where a cross passage meets a running tunnel, the step-plate enlargements where a tunnel changes size. And because the excavation is broken into parts, very large openings can be built as a sequence of smaller ones, each stable on its own. The method's applicability is defined by the ground's stand-up time. Each excavated section must remain stable for long enough to place the sprayed concrete and for that concrete to gain enough strength to carry load. In stiff clays such as London Clay, the stand-up time is long enough that open-face sprayed concrete tunneling is routine, and the method has been used on most of the station caverns and passages of major London projects of the past three decades, including the Jubilee Line Extension and the Elizabeth line. In granular soils below the water table, stand-up time is effectively zero, and the method can only be used after the ground has been treated by dewatering, grouting or freezing, which adds cost and time and creates its own risks. Table 2 compares the three methods across the characteristics that typically decide the choice. Table 2. Comparison of the main soft-ground tunneling methods. Characteristic EPB machine Slurry machine Sprayed concrete (NATM/SCL) Face support Conditioned spoil under pressure Pressurized bentonite slurry Ground's own stand-up time, plus face support measures Best-suited soils Clays, silts, silty sands Sands, gravels, high-permeability ground Stiff clays, weak rock, treated ground Water pressure tolerance Moderate; screw sealing limits it High; well suited to subaqueous work Low unless ground is treated Cross-section Circular, fixed Circular, fixed Any shape, variable Economic length Long drives, typically over 1–2 km Long drives Short tunnels, caverns, junctions Surface plant Spoil handling, conditioning agents Slurry treatment plant, large footprint Concrete batching, modest footprint Source: compiled from ITA-AITES and BTS guidance and general practice. The economics of the choice The choice is rarely made on ground conditions alone. A tunnel boring machine imposes a set of fixed costs: the machine itself, launch and reception shafts large enough to assemble and remove it, a segment factory or supply contract, conveyor systems, power supply at the scale of several megawatts, and in the case of slurry, a separation plant occupying a substantial area of urban land. Against those fixed costs, the machine offers high and predictable advance rates. A well-run EPB drive in suitable ground commonly averages fifteen to twenty meters per day over its length, and peak days considerably higher. A sprayed concrete heading of comparable size might advance a few meters per day. The machine also offers a different risk profile. It eliminates most of the risk of face collapse, but concentrates risk in the machine itself: a breakdown, a stuck shield, a seized main bearing or a wrong choice of machine type can halt progress for months. The Alaskan Way Viaduct replacement in Seattle offers the most widely cited example. The 17.5-meter EPB machine, nicknamed Bertha and at the time the largest-diameter tunneling machine in the world, stopped in December 2013 after only about a thousand feet of progress, with damage to its main bearing seals. Recovering it required excavating a rescue shaft more than one hundred feet deep in front of the machine, lifting out and repairing the cutterhead, and restarting the drive in December 2015. The tunnel eventually opened to traffic in February 2019, years later than planned, and the resulting disputes over who would bear the cost went through extended litigation. The episode illustrates that a closed-face machine replaces the many small risks of open-face excavation with a smaller number of very large ones. Sprayed concrete methods have the opposite profile. They require little capital investment and can adapt to changing conditions, but every round of excavation creates an opportunity for failure, and safety depends on the continuous competence of the workforce and supervision. The Heathrow Express collapse of October 1994, discussed in Chapter 6, showed how a sequence of small departures from good practice in a sprayed concrete tunnel can accumulate into a major collapse. Deciding on the interfaces On most large urban projects, the real design question is not which method to use, but where each method should end and the next begin. A typical metro line uses EPB or slurry machines for the running tunnels between stations and sprayed concrete methods, or cut-and-cover excavations, for the stations, cross passages, ventilation adits and turnouts. The interfaces between the methods require their own planning. Where a sprayed concrete passage is to be opened from a bored tunnel, the segmental lining must be designed with openings that can be cut, often with special steel or fiber-reinforced segments and temporary propping. Where a machine must pass through a station that has already been built, the station must be designed to receive it and to allow it to be relaunched. Where a machine must break into or out of a shaft, the ground at the shaft wall must be treated so that it stands while the shaft wall is opened and before the machine's face seals against the ground. These interfaces are often where the most serious incidents occur. The collapse of a cross passage under construction on Shanghai Metro Line 4 in July 2003, discussed in Chapter 8, happened at exactly such an interface: a hand-excavated passage between two bored tunnels in ground stabilized by freezing, in a confined aquifer under the Huangpu River. The bored tunnels were not the weak point. The connection between them was. The method choice, then, is not a single decision but a map of decisions along the alignment, each one tied to the ground at that point, the structure above it, the water in it, and the method used on either side. Getting that map right does not guarantee success. Getting it wrong almost guarantees difficulty. With the map drawn, the work of controlling the ground begins, and for most of a modern tunnel's length, that work is done by an EPB machine holding the face. Chapter 3. Holding the Face: How an Earth Pressure Balance Machine Works An Earth Pressure Balance machine is, at bottom, a device for keeping a moving hole full. The excavation chamber behind the cutterhead is packed with the soil that has just been cut, and that soil is held under pressure so that it pushes back against the ground in front of it with roughly the same force the ground exerts on it. If the chamber pressure is right, the ground at the face neither moves into the machine nor is pushed away from it, and the surface above barely notices the tunnel passing. If it is too low, the ground moves in, and settlement follows, perhaps a sinkhole. If it is too high, the ground is pushed upward and outward, causing heave or blowing pressurized spoil and conditioning agents up through the ground to the surface. The operator's job, minute by minute, is to keep the pressure within that window while the machine advances and the soil flows through it. The anatomy of the machine Moving from front to back, an EPB machine consists of the cutterhead, the excavation chamber, the pressure bulkhead, the shield with its thrust cylinders, the erector and the tail, followed by a long train of trailing gantries known as the backup. The cutterhead is a rotating steel disc or spoke structure, as large as the tunnel's excavated diameter, carrying cutting tools on its front face. In soft ground these are mostly drag tools, also called scrapers or rippers, which cut soil by shearing it as the head rotates, and their layout determines how evenly the face is excavated. Where rock or boulders are expected, disc cutters are fitted, which crush rock by rolling under very high load. Many machines carry both. The proportion of the cutterhead face that is open, through which soil passes into the chamber, is called the opening ratio, and it is an important design parameter. A high opening ratio, perhaps forty percent or more, lets soil flow freely into the chamber and transmits chamber pressure directly to the face, reducing the risk that soil will compact and clog in the cutterhead. A low opening ratio gives more mechanical support to the face, but tends to clog in sticky clays. Behind the cutterhead is the excavation chamber, also called the working chamber or plenum. It is bounded at the rear by the pressure bulkhead, a thick steel wall that separates the pressurized chamber from the rest of the machine, which is at atmospheric pressure. Through the bulkhead pass the main drive, the screw conveyor, injection ports for conditioning agents, pressure sensors at several heights, and a personnel airlock through which workers can enter the chamber under pressure when tools must be changed or obstructions removed. The screw conveyor is the key to pressure control. It is an inclined auger, typically one meter or so in diameter, whose lower end opens into the bottom of the chamber and whose upper end discharges onto a conveyor belt inside the machine. As the screw turns, it draws soil out of the chamber. The soil inside the screw casing forms a plug, and as the soil moves up the screw it loses pressure progressively, so that it emerges at the discharge gate at close to atmospheric pressure. The screw thus acts as a pressure-reducing seal. Its effectiveness depends entirely on the soil inside it being able to form such a plug: plastic, of low permeability, and able to sustain a pressure gradient along its length. A gate at the discharge end can be partly closed to increase the resistance and support the plug, and many machines have a second screw or a piston-type discharger for use in difficult conditions. The shield is the steel cylinder that protects the machine and the workers inside it. Arranged around its inner circumference are the thrust cylinders, hydraulic jacks that push against the most recently erected lining ring to drive the machine forward. A typical metro-sized machine has around thirty thrust cylinders arranged in groups, and the total thrust available can reach tens of thousands of kilonewtons. By applying more pressure to some groups than others, the operator steers the machine. Many modern machines also have an articulation joint partway along the shield, which allows the front section to be angled relative to the rear to negotiate curves. At the rear of the shield, inside the tail, the segment erector picks up precast concrete segments and places them to form a ring. A ring of a metro tunnel typically consists of five to eight segments plus a smaller tapered key segment, each segment perhaps 1.2 to 2 meters wide along the tunnel axis and 250 to 400 millimeters thick. The rear of the shield carries the tail seal, usually several rows of steel wire brushes packed with a dense grease, which prevents groundwater and grout from entering the machine through the gap between the shield and the lining. Behind the shield, the backup gantries carry the hydraulic power packs, transformers, grout tanks and pumps, conditioning agent equipment, the segment feeder, ventilation fans and the control cabin. The total length of machine and backup can exceed one hundred meters. Defining the target pressure How much pressure should the chamber hold? The answer starts with the pressure the ground and water exert at the face. At any given level on the face, the water pressure is the hydrostatic pressure from the groundwater table above. The earth pressure is more complicated, because it depends on how much the soil at the face is allowed to deform. If the face is held rigidly in place, it carries the earth pressure at rest. If it is allowed to relax slightly, the soil mobilizes some of its strength and the pressure needed to hold it falls toward a lower value called the active pressure. If the face is pushed into the ground, the pressure rises toward a much higher passive value, at which point the ground in front of the machine is driven upward or sideways. A rational target pressure must therefore lie above the minimum needed for stability and well below the pressure that would cause blowout or heave. In practice, designers calculate a lower bound, which is the water pressure plus an earth pressure sufficient to prevent face collapse, plus a margin to account for fluctuations, and an upper bound based on the weight of the overburden above the crown. The upper bound recognizes that, if the support pressure at the crown exceeds the total vertical stress of the soil above, nothing holds the soil down: pressurized spoil or air can lift it and break through to the surface. Between these bounds, the designer chooses a target pressure for each section of the drive, adjusted for changes in cover, water table and soil type. Several calculation methods are used for the stability bound. The most widespread in European practice is the wedge-and-silo model associated with Horn, developed in the 1960s and refined since by Anagnostou and Kovári in the 1990s and by others. It considers a wedge-shaped failure body in front of the face, loaded from above by a prism of soil reaching up to the surface, with arching within the prism reducing the vertical load. The support pressure needed to keep the wedge in equilibrium can then be calculated. The German Tunnelling Committee, DAUB, has published recommendations based on this approach, and they are widely used in Europe and in projects designed by European engineers elsewhere. Other designers use limit analysis methods or numerical modeling. All of them agree on the general features: the required pressure increases with the face diameter, decreases with soil strength, and is dominated below the water table by the water pressure, which must be balanced in full. Water pressure is the key. In a sandy soil below the water table, the effective support must act on the soil skeleton, and in an EPB machine that means the conditioned spoil in the chamber must be sufficiently impermeable that the water pressure at the face is balanced by pressure in the spoil rather than simply flowing through it. If the spoil in the chamber is permeable, water from the face flows into the chamber, the effective stress in the spoil falls, and the pressure that reaches the soil skeleton at the face may be much less than the gauge reading suggests. This is one of the main reasons soil conditioning matters so much, and it is taken up in the next chapter. The pressure profile across the face A chamber full of soil does not have a single pressure. Because the spoil has weight, pressure increases with depth across the chamber, and because the soil is a paste rather than a liquid, its pressure gradient depends on its density and its internal friction. A typical EPB machine has several pressure sensors on the bulkhead at different heights. The crown sensor reads the lowest value and is usually the one that is controlled, because the crown is where face pressure is most critical for settlement and where a deficit leads soonest to collapse toward the surface. The vertical pressure gradient in the chamber tells the operator a good deal about the state of the spoil. If the gradient corresponds to the unit weight of a well-conditioned paste, the chamber is full and the spoil is behaving uniformly. If the gradient is much lower than expected, the upper part of the chamber may contain air or foam rather than spoil, so the chamber is only partially full, and the crown is supported by gas pressure alone. This is a dangerous condition, because gas can escape through the ground quickly, and the face pressure can drop sharply. If the gradient is very high, the spoil may be compacted or contain too much water. Controlling the pressure in real time In operation, the chamber pressure is controlled by balancing two flows: the volume of soil entering the chamber as the machine advances, and the volume leaving it through the screw. The volume entering depends on the advance rate and the face area, adjusted for bulking of the soil as it is cut and for added conditioning agents. The volume leaving depends on screw speed, screw diameter, the degree of filling of the screw flights, and the discharge gate opening. If the operator advances faster without speeding up the screw, more material enters than leaves and the pressure rises. If the screw is speeded up without increasing advance, the pressure falls. Modern machines can control this balance automatically, adjusting screw speed to hold a set pressure while the operator controls advance and steering. Even so, experienced operators watch the pressure trace continuously. Fluctuations of a few tenths of a bar during a single advance are normal, but abrupt drops or sustained trends indicate that something is wrong, such as an unexpected change in the soil, a blocked conditioning line, a screw plug that has failed, or a void ahead of the face. The greatest vulnerability in EPB pressure control occurs during ring building. When the machine has advanced one ring width, it stops, and the thrust cylinders are retracted in groups to make room for each new segment. While the machine is stationary, nothing feeds material into the chamber, and spoil can consolidate, water can drain out through the face or the screw, and the pressure can decay. Operators typically counter this by closing the screw gate and, if necessary, injecting additional conditioning agent or bentonite to maintain pressure. If the machine must stop for longer, for example at the end of a shift, for maintenance or because of a failure on the backup, the risk of pressure loss increases with the duration of the stop. Many settlement incidents have occurred not during excavation but during a prolonged stop, when the pressure slowly drained away unnoticed. Monitoring the spoil Pressure alone does not tell the operator whether the face is stable. A machine can maintain the target pressure in the chamber while removing more soil than it has excavated, if the ground ahead is running into the chamber as fast as the screw removes it. That is called overexcavation, and it is the direct mechanism by which volume loss and sinkholes develop above EPB machines. The second essential control is therefore a check on the quantity of soil removed. The most common method is to weigh the spoil continuously, using belt scales on the conveyor that carries spoil from the screw, and to compare the weight per ring with the theoretical weight of the excavated volume. More sophisticated systems combine belt scales with laser scanners that measure the volume of material on the belt, allowing density to be calculated. The theoretical figure must account for the weight of conditioning agents and any water added, and the in-situ density of the soil, which varies. Even with these corrections, the measurements are imprecise, typically within a few percent, and a single ring's deviation means little. A consistent excess over several rings, however, means the machine is removing more ground than its own volume, and that ground must be coming from somewhere above or around it. Experienced crews also look at the spoil itself. A change in color, texture or moisture content often signals a change in geology before it appears in the pressure trace. Spoil that comes off the belt too wet indicates excess water in the chamber; spoil that is too dry and crumbly indicates inadequate conditioning; the appearance of sand in what should be clay suggests a sand lens at the face. Many tunnels have a spoil inspector whose job is to sample the spoil each ring and record its properties. Settlement, and what it tells you The ultimate test of face control is what happens at the surface and in the buildings above. On urban projects, the ground above the alignment is instrumented densely before the machine arrives: settlement markers set in the road surface, monitoring points on building facades, inclinometers and extensometers in boreholes to measure movement at depth, and, on modern projects, automated total stations that survey hundreds of prisms continuously and report results in near real time. Satellite radar interferometry, known as InSAR, has become a common supplementary tool, able to measure millimeter-scale surface movements over wide areas from repeated satellite passes. The monitoring data feed back into machine operation through a system of trigger levels. Typically, designers set a series of thresholds. At the first, often called green or alert, the monitoring frequency is increased and the data are reviewed. At the second, amber or action, the engineering team meets to review the machine parameters and may adjust face pressure, grout volume or advance rate. At the third, red or alarm, the contingency plan is invoked, which may include stopping the machine, installing additional protective measures, or evacuating buildings. The trigger levels are derived from the predicted settlement and the tolerance of each structure, so a historic building may have much tighter limits than an open car park. This feedback loop is the practical expression of control. The face pressure is set from design calculations, adjusted from the pressure and spoil weight data in the machine, and verified against the ground movement at the surface. When the surface data show more settlement than predicted, the operators increase pressure or tighten the spoil balance. When the data show heave, they reduce pressure. Over the course of a long drive, the team builds up an understanding of how the machine and the ground respond to each other, and the parameters are refined continuously. Where face control fails EPB face control fails most often in four situations. The first is a change in ground that the machine does not detect in time, particularly a transition from clay into sand or gravel, where the conditioning regime suited to clay is inadequate for the new material and the screw plug fails. The second is a prolonged stop with inadequate pressure maintenance. The third is a mixed face, especially where a hard layer or a boulder sits at the bottom of the face and soft ground above it, so that the machine advances slowly while soft ground runs from above. The fourth is shallow cover, where the margin between the minimum and maximum pressures narrows to almost nothing, and any fluctuation either lets the ground in or blows out to the surface. Under rivers, the shallow-cover problem combines with high water pressure and permeable riverbed deposits, making it one of the most delicate conditions an EPB machine can face. In every one of these cases, the underlying difficulty is the same: the soil in the chamber is not behaving as a uniform, plastic, low-permeability paste that can transmit pressure to the face and seal the screw. The machine's control depends on the properties of the material it is digging. That makes the conditioning of that material, the subject of the next chapter, as important to EPB tunneling as the steel of the machine itself. Hashtags: #TunnelingAndUndergroundSpaceEngineering #SoftGroundTunneling #TunnelBoringMachines #EarthPressureBalance #EPBTunneling #ShotcreteLining #SprayedConcreteLining #NewAustrianTunnelingMethod #GroundSupportInteraction #FacePressureControl #GroundwaterControl #EffectiveStress #GroundSettlement #VolumeLoss #SoilConditioning #ScrewConveyor #SegmentalLining #TailVoidGrouting #TunnelFaceStability #StandUpTime #HyperbaricIntervention #CrossPassages #ObservationalMethod #TunnelRiskManagement #FutureOfUndergroundEngineering

Download the Book (PDF): Introduction In February 2013 a nine-year-old girl named Ella Adoo-Kissi-Debrah died in London after a final, catastrophic asthma attack. She had lived for most of her short life about 25 metres from the South Circular Road in Lewisham, one of the busiest arterial routes in the south of the city. Over the preceding three years she had been admitted to hospital dozens of times. Her mother, a former teacher, spent the following years pressing for a second inquest, and in December 2020 a coroner concluded that air pollution had made a material contribution to Ella's death. It was the first time in the United Kingdom that air pollution had been named on a death certificate. The coroner also found that nitrogen dioxide levels near her home had exceeded the legal limits throughout the period of her illness, and that the failure to reduce them had possibly contributed to her death. Clinicians who cared for Ella would have recognised everything about her disease except its cause. They would have known the drugs, the peak-flow diaries, the admission criteria, the escalation pathways. What sat outside the frame of any single consultation was the road. The road had been put there by decisions about traffic, land use, housing and freight made over decades by people who had never met her. Its exhaust was a dose she received every night, and no inhaler could subtract it. This booklet is about that road, and about the parks, trees, supermarkets, pavements and zoning codes that sit alongside it. Its argument is simple to state and harder to act on: the built environment is an exposure, in the strict epidemiological sense, and it should be treated with the same seriousness that medicine brings to drugs, toxins and infectious agents. The shape of a city delivers doses of heat, pollutants, physical activity and food to the people who live in it. Those doses are measurable. They are distributed unequally, and the inequality is not an accident of geography but the legible result of policy, much of it written in the twentieth century and still in force in concrete and asphalt. And because the doses were designed, they can be redesigned. Why this matters to clinicians There is a reasonable objection that urban design belongs to planners and engineers, not to people who work in clinics and health departments. The objection does not survive contact with the evidence. The children of the Southern California Children's Health Study who lived within 500 metres of a freeway reached the end of adolescence with measurably smaller lungs than those who lived 1,500 metres or more away. The deficit was independent of regional air quality, which means that the specific geometry of where a family lived mattered on top of the general pollution of the region. A deficit in lung growth at eighteen is not recoverable; it is a starting point for the rest of adult life. The same pattern recurs across the topics in this booklet. During the Chicago heat wave of July 1995, more than 700 people died in excess of what would have been expected. The case-control study that followed found that living alone, being confined to bed and living on a top floor carried large excess risks, and that a working air conditioner and access to transport were strongly protective. The deaths were distributed not randomly across the city but by housing type, social isolation and neighbourhood. Heat is a hazard that buildings, streets and trees either amplify or blunt. These are clinical outcomes. Asthma admissions, heat stroke, cardiovascular deaths in heat waves, type 2 diabetes and obesity all have environmental gradients that are as steep as many of the risk factors medicine routinely screens for. A clinician who asks about smoking but not about the distance between a child's bedroom and a motorway is taking an incomplete history. What the booklet argues The booklet makes four connected claims. The first is that the built environment can be studied as an exposure, with real metrics, real dose-response relationships and real methodological pitfalls. Chapter 1 sets out how this is done and why the hardest problem in the field is not measurement but self-selection: people choose where to live, and the same things that drive that choice also drive their health. The second is that the environmental burdens of present-day cities were largely distributed by policy. Chapter 2 follows the line from early zoning law, mortgage-risk mapping in the 1930s and the urban highway programmes of the 1950s and 1960s to the asthma, heat and food-access maps of today. The point is not historical grievance; it is causal. If the exposures were placed by decision, they are amenable to decision. The third is that three exposures in particular carry strong and specific evidence. Chapter 3 examines walkability and its relationship to physical activity, obesity and diabetes. Chapter 4 examines tree canopy and the urban heat island. Chapter 5 examines food access and the zoning laws that shape it, a topic on which the evidence turned out to be considerably more complicated than early advocates expected. Chapter 6 examines the near-road environment and children's lungs, which is probably the most robust single body of evidence in the field. The fourth is that interventions on the built environment can change health outcomes, and that we now have enough natural experiments to say something about which ones work, how quickly, and for whom. Chapter 7 reviews that evidence, from the Atlanta Olympics of 1996 to low-emission zones and superblocks. Chapter 8 turns to what clinicians and public-health practitioners can do, from the environmental history taken in a clinic room to the testimony given at a zoning hearing. What the booklet does not do A short book has to choose. This one does not attempt a general survey of urban health, which would need to cover housing quality, indoor air, noise, violence, transport injury, water and sanitation, and much else besides. Some of these appear where they bear directly on the main argument; indoor allergens, for example, are unavoidable in any honest discussion of childhood asthma, and traffic injury cannot be separated from the case for walkable streets. But the focus throughout is on the three themes named in the title and the description: walkability, green space and heat, and the respiratory consequences of living near traffic, together with the zoning decisions that shape food environments. Much of the evidence comes from the United States and Europe, because that is where most of the large cohorts, natural experiments and health impact assessments have been done. The mechanisms are general. Lima, Shanghai and Bogotá rank among the cities with the highest estimated burdens of new childhood asthma attributable to traffic-related nitrogen dioxide, and rapid urbanisation in Africa and South Asia is laying down the street networks and land-use patterns that will govern exposure for the rest of this century. The lessons of the older industrial cities are most valuable where the concrete has not yet set. Finally, the booklet tries throughout to separate what is known from what is plausible. The field has its share of enthusiastic cross-sectional findings that have not survived better designs. Some of the most widely promoted interventions, such as placing supermarkets in food deserts, produced smaller effects than expected when they were evaluated properly. Saying so is not an argument for inaction. It is an argument for spending effort where the evidence is strongest, and for building evaluation into the interventions that cities undertake, so that the next generation of decisions rests on better ground than this one. A note on clinical content Where the booklet touches on clinical management, as it does for asthma and heat illness in Chapter 8, it does so to show how environmental knowledge changes clinical practice, not to serve as a treatment guide. Thresholds, guideline recommendations and air-quality standards are stated as they stand in mainstream guidance at the time of writing; readers should defer to current national guidelines and local protocols for individual patients. Chapter 1: The City as Exposure Epidemiology is good at exposures that come in bottles, packets and syringes. A drug has a dose, a route and a duration. A cigarette can be counted. Even an infectious agent, invisible as it is, can be cultured and typed. The built environment fits none of these templates comfortably. Nobody swallows a street network. A park is not taken twice daily. And yet the city behaves, for the people who live in it, very much like a chronic exposure: continuous, cumulative, largely involuntary once a home has been chosen, and delivered at different intensities to different people according to where they sleep, work and send their children to school. This chapter sets out how the field turns the shape of a city into something that can be measured and related to health, and why the most important methodological question is not how to measure the environment but how to deal with the fact that people are not randomly assigned to it. From place to dose The first move is to recognise that the built environment acts on health through a small number of identifiable pathways, each of which has its own dose metric. Most of the effects discussed in this booklet run through four of them. The first is physical activity. The arrangement of homes, shops, schools and workplaces, and the streets that connect them, determines how far people must travel for everyday purposes and whether walking or cycling is a reasonable way to make the trip. A neighbourhood in which the nearest grocery store is a 25-minute drive along an arterial road with no pavement produces a different weekly dose of moderate activity from one in which the same store is a seven-minute walk along a tree-lined street. The dose here is minutes of moderate-to-vigorous activity, and it can be measured with accelerometers rather than recalled. The second is air pollution, particularly the mixture emitted by motor traffic. Its dose depends on emission density, which is a function of traffic volume, fleet composition and congestion, and on distance and dispersion. Some traffic pollutants, including ultrafine particles, black carbon and freshly emitted nitrogen oxides, fall off steeply within a few hundred metres of a major road. Others, such as ozone and fine particulate matter, form or mix over a whole region. The distinction matters, because it means that a child's exposure is shaped both by the region in which she lives and by the particular block. The third is heat. Dark, impervious surfaces absorb solar radiation and release it slowly; buildings trap heat and reduce airflow; vehicles and air conditioners add waste heat; the absence of vegetation removes the cooling of shade and evapotranspiration. The result is the urban heat island, which is not one island but an archipelago of hotter and cooler patches within a city, sometimes differing by several degrees between neighbourhoods a few kilometres apart. The fourth is food access, which is the most contested of the four. The environment shapes which foods are physically near, which are cheap, which are advertised and which are convenient. Whether physical proximity to a supermarket actually changes diet is, as Chapter 5 will show, a much harder question than it first appeared. Around these four sit others: noise, which has its own cardiovascular literature; the psychological effects of green space and of visible disorder; traffic injury, which is the most direct and least ambiguous way that street design kills people; and social contact, which the Chicago heat wave showed to be a life-or-death variable. These will appear where they bear on the main argument. Each pathway needs a metric that can be attached to a person or a place. The main ones in current use, and what each actually captures, are set out in Table 1. The point of laying them side by side is that each metric carries assumptions, and those assumptions often decide whether a study finds anything at all. Table 1. Common metrics of built-environment exposure and what they measure. Pathway Typical metric What it captures Main limitation Physical activity Walkability index Potential for walking trips Opportunity, not behaviour Air pollution Road distance; modelled NO2 or PM2.5 Outdoor level at home Ignores school, work, transit Heat Satellite land surface temperature Surface heat at a point Not the air people breathe Green space Vegetation index; canopy percentage Amount of vegetation Greenness is not usability Food access Distance to nearest supermarket Proximity to food retail Few shop at the nearest store The buffer problem and the moving person Nearly every metric in Table 1 is computed within a buffer, a circle or street-network distance drawn around a home address. Choose 400 metres and you capture what a person might walk to; choose 1,600 metres and you capture a neighbourhood. Studies routinely report associations at several buffer sizes, and the results do not always agree. This is not merely a statistical nuisance. It reflects a real question about the spatial scale at which an environmental feature acts. A park 300 metres away might be used every day by a toddler and her parent; one 1.5 kilometres away might never be visited. A highway 300 metres away matters for ultrafine particles; one 1.5 kilometres away barely matters for them but may still matter for regional ozone. Geographers call the general version of this the modifiable areal unit problem: the answer depends on how the map is cut. A related problem, sometimes called the uncertain geographic context problem, is that the relevant area is not fixed at all. People move through the city. A child who lives on a quiet street may attend a school beside a motorway and spend forty minutes a day in a car on that motorway. An adult may live in a walkable inner-city flat but work in an office park reachable only by car. The cost of ignoring this is not trivial. In the Cincinnati Childhood Allergy and Air Pollution Study, which followed children from birth to age seven, more than half the cohort moved at least once. Each move tended to take families to addresses with somewhat less traffic pollution and somewhat more greenery. Using only the birth address, or only the latest address, rather than the full residential history produced exposure misclassification that biased the association between traffic pollution and asthma toward the null. In plain terms, simpler exposure assignment made the real effect look smaller than it was. The same logic explains why the California Children's Health Study, discussed at length in Chapter 6, found separate effects for traffic near home and traffic near school. A child's lungs do not know where her address is recorded; they integrate whatever air they breathe over the course of the day. Advances in personal sensors, GPS tracking and high-resolution spatiotemporal models are slowly making it possible to measure this integrated dose. Most of the evidence in this booklet, however, still rests on residential address, and it should be read with the expectation that measurement error has usually made the associations smaller, not larger. Self-selection: the central difficulty If measurement error mostly biases toward the null, the next problem can bias in either direction and is far more serious. People choose where to live, and the reasons that drive the choice are often the same things that drive health. Consider walkability. A young professional who enjoys walking, does not want to own a car and already exercises regularly may deliberately rent a flat in a dense, mixed-use district. If researchers then find that residents of walkable districts are slimmer and more active, how much of the difference is the district and how much is the kind of person who chose it? Planners call this residential self-selection. Epidemiologists recognise it as confounding by preference, and it is extremely difficult to measure because preferences are rarely recorded. Self-selection runs the other way too, and here it intersects with poverty. Many of the environments with the worst exposures, such as housing alongside motorways, heavy industry or freight corridors, are cheap precisely because they are unpleasant. Families end up there not by preference but by constraint. Their health may be worse for reasons that have nothing to do with the motorway: lower income, more stressful work, more crowded housing, less access to care. A naive comparison would attribute all of this to the road. The field has developed several strategies against self-selection, each with characteristic strengths. Adjusting for attitudes and preferences. Some studies ask residents directly about their reasons for choosing a neighbourhood and their attitudes to walking or driving, then adjust for these. The International Physical Activity and Environment Network (IPEN) adult study went further, deliberately sampling residents from neighbourhoods that varied independently in walkability and socioeconomic status across fourteen cities in ten countries, so that the two would not be confounded by design. Such approaches reduce bias but cannot eliminate it, because no questionnaire captures every relevant preference. Longitudinal designs. Following people over time, especially before and after they move, allows each person to serve as his or her own control. If someone's walking increases after moving to a more walkable area, and the increase is proportionate to the change in walkability, the case for a causal effect strengthens. Such studies consistently find smaller effects than cross-sectional ones, which is exactly what one would expect if cross-sectional studies were partly measuring selection. Natural experiments. Sometimes the environment changes around people who did not choose the change. A highway is built or removed; a pest kills millions of trees; an Olympic Games reroutes a city's traffic for seventeen days; a supermarket opens in one neighbourhood but not another that is otherwise similar. These events break the link between personal preference and exposure. Chapters 4, 5 and 7 lean heavily on them. Randomisation. Occasionally, true randomisation of the environment is possible. The Moving to Opportunity experiment in the United States, run by the Department of Housing and Urban Development from 1994 to 1998, randomly assigned 4,498 women with children living in public housing in high-poverty neighbourhoods to one of three groups. One group received housing vouchers usable only in low-poverty census tracts, with counselling on moving; one received unrestricted vouchers; one received neither. When participants were followed up a decade or more later, those offered the low-poverty voucher had lower prevalence of extreme obesity (a body-mass index of 35 or more, and of 40 or more) and of diabetes, defined by glycated haemoglobin of 6.5 percent or more, than controls. The absolute differences were around three to five percentage points. The mechanisms remain unclear, because a low-poverty neighbourhood differs from a high-poverty one in countless ways at once. But the experiment established something important: neighbourhood conditions can cause measurable changes in cardiometabolic disease, not merely accompany them. Cities cannot, of course, be randomised wholesale. But the vacant-lot greening trial in Philadelphia, described in Chapter 4, shows that pieces of them can be, and that such trials are feasible when researchers and city agencies work together. Ecological fallacy and the meaning of an area-level finding Much of the evidence in urban health is ecological: it relates characteristics of areas, such as census tracts or counties, to rates of disease in those areas. Ecological studies are cheap, often use complete administrative data and can reveal patterns that individual-level studies miss. Their weakness is well known. An association between areas need not hold for individuals within them. If tracts with more street trees have lower childhood asthma prevalence, it does not follow that the particular children who live under the trees are the ones who are spared. The authors of one of the best-known studies of this kind were explicit about the limitation. Working with data from 42 health service catchment areas in New York City, they found that an increase in street-tree density of one standard deviation, about 343 trees per square kilometre, was associated with lower prevalence of asthma among children aged four and five, with a relative risk of 0.71. They stated plainly that the design did not permit the inference that trees cause less asthma at the individual level. The value of the finding was to point at a hypothesis worth testing with stronger designs. There is a subtler point here that the clinical reader should hold on to. For some questions, the area-level association is the policy-relevant one. A city deciding whether to plant trees across a district is intervening at the area level. If greening a district reduces its asthma rate, it matters little to the planner which specific children benefited. The ecological fallacy is a fallacy only when one tries to make individual-level causal claims from area-level data. Where the intervention itself is area-level, the area-level estimate is the right target, provided that confounding at the area level has been dealt with. Effect sizes, populations and why small effects matter Built-environment effects are usually modest for any one person. A child near a freeway loses tens of millilitres of lung growth rather than litres. A resident of a highly walkable neighbourhood may get an extra hour or so of physical activity a week. A city-wide tree-planting programme might lower summer temperatures by a fraction of a degree. Clinicians trained to think about the individual patient in front of them may be tempted to dismiss effects of this size. That would be a mistake, for a reason Geoffrey Rose articulated in the 1980s: a small shift in the distribution of a risk factor across an entire population can prevent more disease than a large change in a small high-risk group. Environmental exposures are the purest example of population-wide risk factors. Everyone in a city breathes its air, experiences its heat and moves through its streets. A modest per-person effect multiplied across millions of people yields large numbers of attributable cases. The estimate that around four million new cases of childhood asthma each year worldwide may be attributable to traffic-related nitrogen dioxide, discussed in Chapter 6, is built from just such modest relative risks applied to very large exposed populations. Two further features make built-environment exposures unusually important. They are long-lasting: a street network laid out in 1960 is still shaping travel behaviour today, and a highway built through a neighbourhood will be emitting pollution beside the same homes for a century. And they are distributed unequally, along lines of income and race that track the policies described in the next chapter. A risk factor that is permanent, population-wide and concentrated among the already disadvantaged is one that public health cannot afford to treat as background. What counts as evidence The rest of the booklet relies on a hierarchy of evidence that is worth making explicit. At the bottom are cross-sectional associations between an environmental metric and a health outcome, adjusted for whatever confounders are available. They are useful for generating hypotheses and for mapping where problems lie, but they are vulnerable to self-selection and to confounding by poverty. Above these are prospective cohort studies that measure exposure before the outcome, follow people over time and, ideally, account for residential moves. The California Children's Health Study is the archetype. These studies establish temporality and allow dose-response relationships to be estimated within individuals. Above these are natural experiments, in which an external event changes the environment for some people and not for others. Their strength depends on how plausibly the change was unrelated to the health of the affected population. At the top are randomised trials of environmental change, which remain rare but are no longer unheard of. Running alongside this hierarchy is mechanistic evidence: controlled exposure studies in which volunteers breathe diesel exhaust in a chamber, animal studies of lung development, physical models of how trees cool a street. Mechanistic evidence cannot by itself establish population effects, but it makes epidemiological associations far more credible when the two agree. Finally there is health impact assessment, which is not evidence of effect at all but a structured way of combining effect estimates with local exposure data to forecast what a policy might achieve. Several of the most-cited numbers in this field, including the estimate that raising tree cover to 30 percent across European cities could prevent thousands of summer deaths, come from health impact assessments. They are valuable for decision-making, but they inherit every uncertainty in the effect estimates they use, and the reader should treat them as projections rather than observations. With these tools in hand, the next question is how the environments being measured came to be the way they are. That turns out to be a story not of geography or chance but of deliberate decisions, many of them written into law. Chapter 2: How Exposure Was Zoned It is tempting to think of a city's environmental hazards as the residue of geography and economics: industry clustered by the river because that was where the barges came, highways ran through the cheapest land because that was efficient, wealthier people moved upwind because they could. There is truth in each of these stories, but taken together they leave out the most important actor. In most cities of the industrialised world, and conspicuously in the United States, the distribution of heat, pollution, trees and food was shaped by law and public policy. Zoning codes decided which land could hold a factory and which could hold only detached houses. Federal mortgage underwriting decided which neighbourhoods would receive investment and which would be starved of it. Highway engineers, using public money and powers of eminent domain, decided whose homes would be demolished to make way for the roads. For a clinical reader this history matters for a precise reason. If present-day exposure gradients were placed by decisions, then they are not fixed features of urban life. They are artefacts, and artefacts can be altered. The history is also the best available explanation for why the exposures examined in the following chapters cluster so tightly together, and why the same children who live nearest the motorway also tend to live in the hottest, least-treed blocks with the fewest grocery stores. The invention of zoning Modern land-use zoning grew out of a public-health impulse. Late nineteenth-century cities mixed tanneries, slaughterhouses, chemical works and tenements on the same streets, and reformers wanted to separate noxious uses from homes. New York City's 1916 zoning resolution, the first comprehensive code in the United States, was driven largely by concern about tall buildings blocking light and air from the streets below. In 1926 the Supreme Court upheld the principle of comprehensive zoning in Village of Euclid v. Ambler Realty Co., and the model it endorsed, known thereafter as Euclidean zoning, spread across the country. Euclidean zoning works by dividing a municipality into districts, each permitted a narrow range of uses arranged in a hierarchy. At the top sits the single-family residential district, the most protected; below it come multifamily housing, commercial uses, light industry and heavy industry. Higher uses are protected from lower ones. In its original form the hierarchy was cumulative, so that houses could be built in industrial districts but factories could not be built in residential ones. That asymmetry had consequences nobody intended in the abstract and many intended in practice. Wealthier, whiter neighbourhoods could secure exclusive residential zoning and keep out industry, apartment buildings and commerce alike. Poorer neighbourhoods, often zoned for industrial or mixed use, received whatever the protected areas excluded. Explicit racial zoning had been struck down by the Supreme Court in Buchanan v. Warley in 1917, but the same effect could be achieved by zoning categories, minimum lot sizes and restrictive covenants. Studies of mid-century zoning maps in many American cities show that Black neighbourhoods were disproportionately zoned for industrial use, which then attracted junkyards, incinerators, rail yards and truck depots. Euclidean zoning also produced a second health consequence, quieter but just as pervasive. By separating homes from shops, schools and workplaces, and by requiring large minimum lot sizes and generous off-street parking, it made the typical post-war suburb nearly impossible to navigate except by car. The walkability problem examined in Chapter 3 is, in large part, a zoning problem. In many American municipalities it remains illegal to build the kind of neighbourhood, with corner shops, flats above stores and houses on small lots, that produces the most walking. The red line In 1933 the federal government created the Home Owners' Loan Corporation to refinance mortgages in default during the Depression. Between 1935 and 1940 the corporation commissioned local real-estate professionals to produce "residential security maps" of some 200 American cities, grading each neighbourhood for mortgage lending risk. Grade A, outlined in green, was "best"; grade B, in blue, "still desirable"; grade C, in yellow, "definitely declining"; and grade D, in red, "hazardous". The appraisers' notes, which survive alongside the maps, make the criteria plain. The presence of Black residents, immigrants or Jewish families was routinely cited as a reason for a low grade. Industrial nuisances, aging housing stock and proximity to rail lines counted too. The maps themselves were one part of a wider apparatus of discriminatory lending, including the underwriting rules of the Federal Housing Administration, which for decades steered insured mortgages away from mixed and Black neighbourhoods. The combined effect was to channel private and public investment toward A and B neighbourhoods and away from C and D ones. Redlined neighbourhoods saw lower home ownership, less maintenance, lower property values and, over time, a lower tax base for the services, including street trees and parks, that local government provides. Redlining has been illegal since the Fair Housing Act of 1968. What makes it valuable to epidemiologists is that the maps were digitised in the 2010s by the University of Richmond's Mapping Inequality project, so that any present-day address can be assigned the grade its neighbourhood received in the 1930s. Because the grades were assigned eighty years ago, they cannot have been caused by present-day health; they are a fixed historical exposure against which modern outcomes can be measured. The results are consistent across exposures. Air pollution. An analysis of 202 American cities found that present-day concentrations of nitrogen dioxide and fine particulate matter rose almost monotonically from grade A to grade D, with nitrogen dioxide levels in D neighbourhoods more than 50 percent higher than in A neighbourhoods. Nationally, the intra-urban disparity by historical grade was larger than the disparity by present-day race and ethnicity. Within each grade, racial disparities persisted, a reminder that redlining was one discriminatory policy among many. Heat. A study of 108 American urban areas found that formerly redlined neighbourhoods were, on average, about 2.6 degrees Celsius hotter in land surface temperature than grade A neighbourhoods in the same cities, with differences of several degrees more in some. The mechanism is straightforward: fewer trees, more pavement, denser and older housing with dark roofs, and more industrial land. Asthma. In eight California cities, census tracts that had been redlined had a median age-adjusted rate of emergency department visits for asthma of 63.5 per 10,000 residents per year between 2011 and 2013, compared with 26.5 in grade A tracts, a ratio of about 2.4. After adjustment for present-day poverty, diesel exhaust emissions and fine particulate matter, redlined tracts still had a relative rate of 1.39. That residual is notable: it suggests that something about the history of these neighbourhoods matters beyond the contemporary measures of pollution and income used in the model, whether it is housing quality, cumulative disinvestment in health services, or exposures that the available metrics capture poorly. These are ecological associations and carry the limitations described in Chapter 1. But they share a feature that makes them unusually persuasive: the exposure was assigned decades before the outcome, by a process whose logic is documented in the appraisers' own words. Few environmental exposures have such clean temporal precedence. Highways through neighbourhoods The Federal-Aid Highway Act of 1956 committed the United States to building a national network of limited-access highways, with the federal government paying 90 percent of the cost. The interstates were conceived as links between cities, but in practice many of the most expensive and consequential miles ran through them, and planners, engineers and local officials frequently chose routes through Black and immigrant neighbourhoods. Sometimes this was justified explicitly as "slum clearance"; sometimes it reflected the political weakness of residents who had been excluded from the mortgage market and from local government; sometimes cheaper land values, themselves a product of redlining, made the route the cheapest option. The examples are well documented. In New Orleans, Interstate 10 was elevated over North Claiborne Avenue, the commercial and social spine of the Tremé neighbourhood, and the avenue's broad median of live oaks was removed. In St. Paul, Interstate 94 cut through Rondo, the city's largest Black neighbourhood, displacing hundreds of families and businesses. In Miami, Interstate 95 and its interchanges sliced through Overtown. In Syracuse, an elevated stretch of Interstate 81 divided the 15th Ward. In the Bronx, the Cross Bronx Expressway, built under Robert Moses, displaced thousands of residents and left the borough with a band of heavy traffic that runs past schools and apartment blocks to this day; the South Bronx has for decades had some of the highest childhood asthma hospitalisation rates in New York City. The direct effects were demolition and displacement. The long-term effects were exposure. The families who remained alongside the new highways, and those who later moved into the cheap housing beside them, received a continuous dose of traffic pollution and noise. Chapter 6 examines in detail what that dose does to children's lungs. The point here is that the near-road population was not randomly assembled. It was created, in many cases, by a sequence of policies that first concentrated disadvantaged families in particular neighbourhoods and then routed the highways through them. The same pattern is visible outside the United States, in different forms. In Britain, inner-city motorway schemes of the 1960s and 1970s, such as Westway in west London, brought heavy traffic to the edges of social-housing estates. In many rapidly growing cities in the global South, informal settlements form along highway margins, rail corridors and industrial zones precisely because that land is unclaimed, cheap or unsafe, and residents there may receive extreme traffic exposures with no formal planning process at all. Ports, freight and the newer geography of diesel The interstates are the most visible example of transport infrastructure placed beside disadvantaged communities, but they are not the only one, and in some places they are no longer the most important. Seaports, rail yards, airports and the distribution warehouses that serve them generate concentrated diesel emissions from ships, locomotives, cargo-handling equipment and, above all, heavy trucks. The ports of Los Angeles and Long Beach, which together form the busiest container port complex in the United States, illustrate the pattern. The communities immediately around them, such as Wilmington, West Long Beach and the neighbourhoods along the Interstate 710 freight corridor, are predominantly Latino and lower-income, and for years recorded some of the highest modelled diesel particulate concentrations and cancer risk estimates in the region. Community campaigns from the early 2000s onward, combined with litigation and state regulation, led to the ports' Clean Air Action Plan of 2006, which required cleaner trucks, shore power for ships at berth and cleaner cargo-handling equipment. Monitored diesel particulate emissions from port operations fell substantially in the following decade. The episode is an early example of exposure reduction driven partly by the people living with the exposure, and it is one strand of the regional air-quality improvement whose effect on children's lungs is examined in Chapter 6. The newer geography of diesel is driven by e-commerce. Large distribution warehouses have proliferated on the edges of metropolitan regions, often on cheap industrial land in lower-income areas, and each can generate hundreds or thousands of heavy truck trips a day. In southern California's Inland Empire, east of Los Angeles, warehouse floor space grew enormously from the 2000s onward. The South Coast Air Quality Management District adopted an Indirect Source Rule in 2021 that requires large warehouses to reduce or offset the emissions of the truck traffic they attract, a regulatory approach that treats the building as responsible for the pollution its use generates. Whether such rules keep pace with logistics growth is an open question. Planning law has historically struggled with these facilities because each one, considered in isolation, may meet every applicable standard. The health burden is cumulative: a neighbourhood next to a freeway, a rail yard and a cluster of warehouses receives the combined plume, but each facility is permitted separately. Several jurisdictions have begun to require cumulative impact assessment. New Jersey's environmental justice law, enacted in 2020, requires the state environmental agency to consider the combined existing burden on designated overburdened communities when deciding permits for certain new or expanded polluting facilities, and allows denial where a facility would add disproportionately to that burden. It is among the first laws in the United States to make cumulative exposure, rather than facility-by-facility compliance, the basis for a permitting decision. How the exposures cluster The policies in this chapter did not deliver one hazard at a time. Zoning, lending and highway routing operated on the same neighbourhoods and reinforced one another. A neighbourhood graded D was likely to be zoned for mixed or industrial use, to lose investment in housing and street trees, to be chosen for a highway route, and later to lose its grocery stores as population and purchasing power declined. The result in many present-day cities is a set of tightly correlated exposures. The hottest blocks tend to be the least treed, the least treed tend to be nearest to heavy traffic, and those blocks also tend to have the fewest full-service food retailers and the most fast-food and convenience outlets. This clustering creates two problems, one analytical and one practical. The analytical problem is confounding between exposures. If heat, pollution, poor food access and low walkability all travel together, a study that measures only one of them will attribute to it the effects of all. Disentangling them requires either measuring them together, which good modern studies try to do, or finding settings where they come apart, which is the work of natural experiments. The practical problem is that interventions focused on a single exposure may disappoint. Planting trees on a street that remains beside a motorway will cool it but leave the pollution largely in place. Opening a supermarket in a neighbourhood where incomes are low and transport is poor may do less than expected, a finding examined in Chapter 5. The history suggests that what was done to these neighbourhoods was done in bundles, and that undoing it may also require bundles. The persistence of built form One more feature of this history matters for anyone hoping to change it. Buildings and street networks are extraordinarily durable. Most of the buildings that will be standing in European and North American cities in 2050 already exist. Street patterns persist for centuries; the medieval layout of the City of London survived the Great Fire of 1666 largely intact. A highway interchange built in 1965 is expected to be maintained, not removed. Zoning codes are amended piecemeal, and amendments that would allow denser or more mixed development frequently meet fierce local opposition from existing residents. This durability cuts both ways. It means that the exposures created in the twentieth century will continue to act unless something is deliberately done about them. It also means that the decisions being made now, in the zoning codes of fast-growing cities and in the routing of new roads, will govern exposure for a very long time. The most cost-effective moment to prevent a near-road school or a treeless housing estate is before it is built. There are signs of change. Several American cities, including Minneapolis in its 2040 comprehensive plan, have moved to end exclusive single-family zoning. A handful of urban highways have been removed or are being removed, among them the Embarcadero Freeway in San Francisco after the 1989 earthquake, the Cheonggyecheon elevated road in Seoul in 2003, and the Interstate 81 viaduct in Syracuse, where construction on a replacement community grid began in 2023. The 2021 federal infrastructure law in the United States created a Reconnecting Communities programme to fund the planning and removal of such barriers. These are modest steps relative to the scale of what was built, but they establish a principle that would have been unthinkable a generation ago: that a road can be judged by its effect on the people beside it, not only by its effect on the people driving on it. With this history in view, the following four chapters take the main exposures one at a time, beginning with the pathway most directly shaped by zoning: whether the neighbourhood makes it easy or hard to walk. Chapter 3: Walkability and the Metabolic City Physical inactivity is one of the leading modifiable causes of non-communicable disease worldwide. The World Health Organization's current guidelines recommend that adults accumulate at least 150 to 300 minutes of moderate-intensity aerobic activity a week, or an equivalent amount of vigorous activity, and that children and adolescents average at least 60 minutes a day. Large shares of adults in high-income countries fall short. Medicine's standard response has been to treat activity as a behaviour to be counselled: advise the patient to exercise, perhaps prescribe a programme, follow up. The results of brief counselling are real but modest, and they decay. The urban-design perspective approaches the problem from the other end. Instead of asking how to persuade people to add exercise to their day, it asks how much activity their day already contains by default, and what in the environment determines it. A person who walks to the bus stop, walks from the station to the office, walks to the shop at lunchtime and walks home from the bus in the evening can accumulate thirty or forty minutes of moderate activity without ever deciding to exercise. A person who drives from an attached garage to a parking structure under the office and back may accumulate almost none. The difference is not willpower. It is built into the street. What walkability is The word walkability is used loosely, but in research it usually refers to a composite index of three or four measurable features of the area around a home, calculated within a street-network buffer of, typically, 500 metres or one kilometre. Residential density is the number of dwellings per unit of residential land. Higher density puts more destinations within walking distance and makes public transport viable, because a bus or train needs enough riders within a short walk of each stop. Street connectivity, usually measured as intersection density, describes how directly the street network links places. A traditional grid with short blocks offers many routes between two points, and the walking distance is close to the straight-line distance. A suburban network of cul-de-sacs and looping collector roads may put a shop 200 metres away as the crow flies but 1.5 kilometres away on foot. Land-use mix measures how many different uses, such as residential, retail, office, institutional and recreational, are present in the area. Without a mix, there is nothing to walk to. Some indices add retail floor-area ratio, which distinguishes shops that front onto the pavement from those set behind large car parks. These features are, almost by definition, the ones that Euclidean zoning, minimum lot sizes and parking requirements suppress. Walkability is the measurable inverse of the post-war suburban template described in Chapter 2. What the indices do not capture is equally important. They say nothing about whether pavements exist or are maintained, whether crossings are safe, whether the street is shaded, whether the walker feels safe from crime, or whether the destinations within reach are ones the resident actually needs. Two neighbourhoods with identical walkability scores can differ greatly in whether anyone walks in them. This is one reason that studies relating walkability to health find consistent associations but rarely large ones. Walkability and activity The most rigorous international evidence comes from the International Physical Activity and Environment Network adult study, which sampled 6,822 adults in fourteen cities across ten countries on five continents, including Bogotá, Cuernavaca, Curitiba, Ghent, Hong Kong, Melbourne, Olomouc, Seattle and Baltimore. Participants were recruited from neighbourhoods selected to vary independently in walkability and socioeconomic status, and their physical activity was measured with accelerometers worn for four to seven days rather than by recall. The study, published in The Lancet in 2016, found that four of six measured environmental attributes were positively and linearly related to moderate-to-vigorous physical activity: net residential density, intersection density, public transport density and the number of parks within walking distance. The difference in activity between residents of the most and least activity-friendly neighbourhoods was between 68 and 89 minutes a week, which is 45 to 59 percent of the 150 minutes recommended by guidelines. The similarity of the associations across very different cities was itself an important result: it suggested that the relationship between urban form and walking is not an artefact of American suburbia but a general property of how people move. A companion analysis of more than 12,000 IPEN participants, focusing on walking and cycling for transport, found positive associations with all four environmental attributes but also found that, except for land-use mix, the relationships were curvilinear, flattening at higher levels. Beyond a certain density, intersection density or number of parks, further increases added little. That is a useful practical finding. It implies that the largest health gains come from moving the least walkable neighbourhoods up to moderate walkability, not from pushing already dense urban cores to greater extremes. Walkability, obesity and diabetes If walkable neighbourhoods produce more walking, they might be expected to produce less obesity and diabetes. The evidence here is consistent in direction and more uncertain in interpretation. Early cross-sectional work in the United States established the pattern. A 2003 analysis of more than 200,000 respondents to the Behavioral Risk Factor Surveillance System across 448 counties constructed a county sprawl index and found that residents of more sprawling counties walked less in their leisure time, weighed more and had a higher prevalence of hypertension than residents of compact counties, after adjustment for demographic and behavioural factors. The associations were small but statistically robust. A study of nearly 11,000 residents of the Atlanta region published the following year found that each quartile increase in land-use mix around a person's home was associated with a 12.2 percent lower likelihood of obesity, each additional hour per day spent in a car with a 6 percent higher likelihood, and each additional kilometre walked per day with a 4.8 percent lower likelihood. The most informative recent study comes from Ontario, where researchers used linked provincial health administrative data covering roughly three million adults aged 30 to 64 a year, living in more than 8,700 neighbourhoods across southern Ontario's cities, between 2001 and 2012. Neighbourhoods were ranked into fifth parts by walkability. In 2001 the adjusted prevalence of overweight and obesity was 43.3 percent in the most walkable fifth and 53.5 percent in the least walkable. Over the following decade, prevalence rose significantly in the three least walkable fifths, by between 5.4 and 9.2 percentage points, but did not change significantly in the two most walkable. Diabetes incidence fell in the two most walkable fifths, from 7.7 to 6.2 per 1,000 in the most walkable, while not changing significantly in the less walkable ones. Residents of the most walkable neighbourhoods walked, cycled and used public transport more and drove less at every time point. The Ontario study is instructive for what it found and what it could not show. Its divergent trends over time are harder to explain away than a single cross-sectional difference, and its population coverage is almost complete. But it was ecological, and its authors noted that leisure-time physical activity, diet and smoking did not vary by walkability, and that walking and cycling had increased only modestly in walkable areas over the period. The mechanism that would connect walkability to slower weight gain was not fully visible in the data. One plausible explanation is that the relevant activity, short incidental walks for errands and transit, is poorly captured by surveys that ask about exercise. Another is that some other feature of walkable neighbourhoods, such as their food environments or their populations, accounts for part of the difference. The study's own conclusion, that further research is needed to establish causality, is the honest one. Self-selection, revisited Walkability is the exposure for which residential self-selection, introduced in Chapter 1, matters most. People who like walking may seek out walkable neighbourhoods; people who prefer driving may seek out suburbs with large garages. Studies that have asked residents about their preferences and adjusted for them consistently find that the association between walkability and walking shrinks but does not disappear. Longitudinal studies of people who move find that walking for transport rises after moving to a more walkable area and falls after moving to a less walkable one, though by less than cross-sectional comparisons would suggest. A reasonable summary of this literature is that walkable environments genuinely increase walking for transport, by an amount that is meaningful at the population level, but that the cross-sectional associations overstate the effect because they partly reflect who chooses to live where. The effect on leisure-time exercise is weaker and less consistent; people who want to go running will find somewhere to run. The effect on body weight is real but small for any individual and plausibly larger at population scale over decades. There is also a selection effect that runs in the opposite direction and is rarely discussed. In many North American cities, the most walkable older neighbourhoods were, for much of the late twentieth century, the poorest, because wealthier households had moved to car-oriented suburbs. In those settings, the healthy-walker selection and the poverty selection pull against each other. The Ontario study observed higher poverty in walkable areas, and still found better trends in weight and diabetes there, which, if anything, strengthens its findings. More recently, gentrification has reversed the pattern in many inner cities, pricing poorer residents out of exactly the neighbourhoods that offer the most incidental activity. Walkability has become, in many markets, a premium good, and the housing affordability problem has become a public-health problem. Safety, traffic and the walking environment A neighbourhood is walkable only if it is safe to walk in. Two kinds of danger dominate. The first is traffic. In the United States, pedestrian deaths have risen substantially since around 2010, reaching their highest levels in four decades in the early 2020s, even as deaths among vehicle occupants rose more slowly. Several factors have been implicated, including the shift in the vehicle fleet toward larger and taller sport-utility vehicles and pickup trucks, whose front ends are more likely to strike the head and torso, and the prevalence of wide, fast arterial roads with long distances between signalised crossings. Those arterial roads are disproportionately located in lower-income areas and the suburbs built for car travel. A walkability index that counts intersections and destinations but ignores five-lane roads with 45 mile-per-hour speed limits may be measuring the potential for walking in a place where walking is, in fact, dangerous. Street design can change this. Narrower lanes, protected crossings, shorter crossing distances, lower speed limits and physical traffic calming all reduce pedestrian injury. The Vision Zero approach, which originated in Sweden in the late 1990s and treats any traffic death as a design failure rather than a matter of individual error, has been adopted by many cities. Its relevance to this chapter is that safe streets are a precondition for the health benefits of walkable ones. The second danger is crime and the fear of it. Parents who believe that the walk to school is unsafe will drive their children, and adults who fear assault will not walk after dark. Perceived safety is partly a function of the built environment itself: lighting, visibility, the presence of other people on the street, and the condition of buildings and vacant land. The Philadelphia vacant-lot experiments described in Chapter 4 show that changes to the physical environment can reduce both crime and fear. Active travel to school Children's activity is a special case. In many high-income countries, the share of children who walk or cycle to school has fallen steeply since the 1960s and 1970s. Several design factors explain much of this: the consolidation of schools into larger campuses farther from homes, the siting of new schools on cheap land at the edges of developments, the absence of safe routes across arterial roads, and the expectation, built into many suburban layouts, that children will be driven. The consequence is twofold. Children lose a daily dose of activity, and the school run adds car trips at precisely the times and places where children congregate, raising pollution and collision risk outside school gates. Some cities have responded with "school streets", in which roads outside schools are closed to motor traffic at opening and closing times. London has implemented several hundred of these since 2020. Early evaluations suggest that they reduce traffic and nitrogen dioxide near school entrances, and that they increase the share of children walking or cycling, although the effect on overall neighbourhood traffic depends on how displaced trips are absorbed. Older adults and the walkable neighbourhood Most walkability research has focused on working-age adults, but the stakes may be highest at the other end of life. For older adults, the neighbourhood is often the whole of the daily environment. Retirement removes the commute; declining vision, balance or confidence may end driving; and the distance to the nearest shop, surgery or bench begins to determine whether a person leaves home at all. The features that matter for older walkers overlap with conventional walkability indices but go further. Even, well-maintained pavements reduce the risk of falls, which are a leading cause of injury, loss of independence and death in later life. Benches at regular intervals allow people with limited exercise tolerance, such as those with heart failure or chronic obstructive pulmonary disease, to make trips in stages. Public toilets, shade, longer pedestrian crossing times at signals and dropped kerbs all extend the range of a walk for someone with reduced mobility. A neighbourhood that scores well on density and land-use mix but has broken pavements, short crossing phases and no seating may be walkable on paper and unwalkable for an eighty-year-old. The connection to the rest of this booklet is direct. The social isolation that proved so lethal in the Chicago heat wave, discussed in Chapter 4, is partly a product of environments that make it hard for older people to get out. A street where an older resident can walk to a shop and back, meeting neighbours on the way, is a street on which someone is more likely to notice when she has not appeared for two days. The World Health Organization's age-friendly cities programme, launched in 2007 and since joined by many hundreds of municipalities, treats outdoor spaces and transport as two of its core domains for precisely this reason. Modelling the gains Because walkability is hard to change quickly and harder to randomise, much of the evidence on what larger-scale changes might achieve comes from modelling. A 2016 series on urban design, transport and health in The Lancet modelled what would happen to health in six cities on four continents, including Melbourne, Boston, London, Copenhagen, Delhi and São Paulo, if land use were made more compact and diverse and travel shifted from private cars toward walking, cycling and public transport. The modelled scenarios produced health gains in every city, with the benefits of increased physical activity generally outweighing the added risks from pollution exposure while walking and cycling and from road injury, and with the largest gains in cardiovascular disease, diabetes and respiratory disease. The size of the gains varied with each city's starting point; car-dependent cities stood to gain most from mode shift, while cities with already high walking rates stood to gain most from making walking and cycling safer and cleaner. Such models depend on assumptions about how much behaviour would actually change, and on dose-response relationships borrowed from other settings. They are not evidence that the changes would occur. What they do show is that, under plausible assumptions, the health case for compact, mixed, walkable cities is robust to the main objections raised against it, namely pollution exposure and injury risk, provided that traffic speeds and volumes are managed alongside density. What walkability does not do on its own The evidence on walkability supports a clear but bounded conclusion. Neighbourhoods with denser, more connected, more mixed and better-served street networks produce more walking for transport, and their residents, on average, have somewhat lower rates of obesity and diabetes, with trends over time that favour them. The effects are modest for individuals and meaningful for populations. Three cautions follow from the rest of this booklet. First, walkability and air pollution can pull in opposite directions. The densest, most connected neighbourhoods often carry the heaviest traffic, and pedestrians and cyclists breathe more air per minute than drivers. Health impact modelling generally finds that the benefits of increased physical activity outweigh the added pollution risk for walkers and cyclists in all but the most polluted cities, but the trade-off is real, and it is strongest along the arterial roads where walking is least pleasant. Walkable neighbourhoods are healthiest when they are also low-traffic neighbourhoods. The superblocks of Barcelona, discussed in Chapter 7, are designed precisely to combine the two. Second, heat matters. A walkable street with no shade in a hot climate may go unused for much of the summer. Chapter 4 turns to the role of trees in making streets usable as well as cooler. Third, the benefits of walkable design accrue to those who live there. If walkable neighbourhoods are scarce and expensive, their health benefits will flow to those who can afford them, and the gradient described in Chapter 2 will be reproduced in a new form. Expanding the supply of walkable neighbourhoods, through zoning reform that permits mixed uses and moderate density across more of the city, is the only durable answer to that problem. Hashtags: #UrbanDesignAndPublicHealth #BuiltEnvironment #Walkability #GreenSpace #ChildhoodAsthma #TrafficRelatedAirPollution #NearRoadExposure #NitrogenDioxide #LungDevelopment #UrbanHeatIsland #TreeCanopy #PhysicalActivity #ActiveTravel #LandUseMix #StreetConnectivity #ResidentialDensity #FoodAccess #EnvironmentalJustice #Redlining #ZoningAndHealth #ResidentialSelfSelection #NaturalExperiments #PopulationHealth #HealthImpactAssessment #FutureOfHealthyUrbanDesign

Download the Book (PDF): Introduction On the evening of 1 September 2021, the remnants of Hurricane Ida passed over New York City and dropped 3.15 inches of rain on Central Park in a single hour. It was the highest hourly total ever recorded there, and it broke a record set only eleven days earlier by Tropical Storm Henri. Water poured down subway stairwells, lifted cars off expressways, and filled basement apartments in Queens to the ceiling. Most of the people who died in the city that night drowned below ground, in their own homes. The sewers did not fail in the sense of breaking. They did exactly what they were built to do, which was to carry a storm of a size that the engineers of an earlier century had judged sufficient. The rain simply arrived faster than the pipes could take it, and the excess went where water always goes: downhill, into the lowest available space. Two months earlier, on 20 July 2021, the Chinese city of Zhengzhou recorded 201.9 millimetres of rain in one hour, close to eight inches. Floodwater broke through a temporary wall at a metro depot, entered the tunnels of Line 5, and trapped passengers in carriages as the water rose to their chests and then their necks. Ten years before that, in July 2011, a cloudburst over Copenhagen delivered well over a hundred millimetres of rain in about two hours in the worst-hit districts and left damage that the city's own later economic analysis put at more than six billion Danish kroner, roughly 800 million euros, before indirect costs were counted. Within months Copenhagen had begun drafting a plan that would change how the city thought about water in its streets. These are not isolated freak events. They are the visible edge of two slow changes that have been converging for decades. The first is that cities have paved over the ground that used to absorb rain. Roofs, roads, car parks and plazas convert a soft, spongy landscape into a hard one, and a hard landscape sheds water quickly. The second is that the atmosphere is warming, and warmer air holds more moisture. The physics of that relationship is old and well understood, and its consequences for short, intense rainfall are now showing up in the records of rain gauges around the world. Put the two together and the problem is plain: more water is falling in shorter bursts onto surfaces that can no longer slow it down. The argument of this book The standard response to urban flooding, for most of the twentieth century, was a bigger pipe. If a neighbourhood flooded, engineers upsized the drains so the water could leave faster. That approach worked well enough when the storms were stable and the cities were smaller. It has two fundamental weaknesses that become obvious once you look at the problem in the right way. The right way to look at it is through the hydrograph: the curve that plots the rate of flow at a point against time. Every storm produces one. A natural catchment produces a low, broad hydrograph, because rain soaks in, travels slowly through soil and vegetation, and arrives at the stream over many hours. A paved catchment produces a tall, narrow hydrograph, because rain runs straight off hard surfaces into gutters and pipes and arrives all at once. Flooding is a problem of the peak of that curve. Pollution from combined sewers is a problem of the volume under it. Both are problems of timing. Once you see the problem as a hydrograph, the weakness of the bigger pipe becomes clear. A bigger pipe does nothing to the shape of the curve. It simply moves the peak downstream faster, often into someone else's neighbourhood or into a river that cannot take it. And because pipes are designed for a particular storm, any storm larger than that one finds the pipe full and spills onto the surface anyway, with no plan for where the surplus will go. This book argues that managing stormwater in a changing climate is fundamentally a problem of shaping the hydrograph, of controlling where water is stored and when it is released, and that the robust answer is a layered system rather than a single solution. Distributed green infrastructure such as permeable pavements and bioswales handles the frequent, small storms that make up most of the annual rainfall and most of the pollution. Large engineered storage, such as deep tunnels and underground basins, captures the volumes that would otherwise overflow from combined sewers. And deliberately designed surface flood routes, roads shaped as channels and parks built to fill, carry the rare extreme storms that no buried system can afford to handle. Each layer attenuates a different part of the hydrograph. None is sufficient on its own. What ties them together, and what makes the whole defensible as engineering rather than aspiration, is hydraulic modelling: the quantitative prediction of how each element changes the timing and magnitude of flow. That last point deserves emphasis because it is where enthusiasm often runs ahead of evidence. Green infrastructure has a powerful appeal. It is visible, it plants trees, it cools streets, and it photographs well. Deep tunnels, by contrast, are invisible, expensive and associated with decades-long legal settlements. It is tempting to frame the choice as a moral one between green and grey. This book does not. A bioswale that infiltrates the first twenty-five millimetres of rain is enormously valuable for water quality and for the frequent storms that cause nuisance flooding, but it will do very little on the night that eight inches fall. A deep tunnel that stores millions of cubic metres will dramatically reduce sewage overflows, but it will not stop surface flooding in a street whose inlets are overwhelmed. The question is never which approach is better. It is which part of the hydrograph each one addresses, how much it costs per unit of attenuation, and how they behave together when the storm exceeds what any of them were designed for. How the book proceeds The chapters build that argument in sequence. The first explains what urbanisation does to the hydrological cycle and why the paved catchment behaves so differently from the natural one. The second turns to the rain itself: how engineers have traditionally described design storms, why the assumption that rainfall statistics are fixed has broken down, and what replaces it. The third sets out the core mathematics of peak flow attenuation, from the rational method to reservoir routing, with a worked example that shows how a storage element flattens and delays a flood peak. The fourth explains how modern hydraulic models represent a whole city's drainage, including the coupling of buried pipes with surface flow, and where those models are trustworthy and where they are not. The next chapters take the engineering elements in turn. Permeable pavements and bioswales each receive a full chapter on design, performance and failure. A further chapter asks the question that the enthusiasm for green infrastructure often skips: does distributed storage actually add up to meaningful attenuation at the scale of a whole catchment, and under what conditions can it make things worse? The book then moves to deep tunnel storage, the largest and most expensive tool, examining Chicago, Milwaukee, London, Washington, Tokyo and Kuala Lumpur. It turns next to the surface itself as a flood conveyance system, the approach pioneered by Copenhagen and Rotterdam, and then to the operation of drainage networks in real time, where sensors and controllable gates can extract more capacity from infrastructure that already exists. The final chapter addresses how to make design decisions when the future rainfall is genuinely uncertain. A few terms recur throughout and are worth fixing at the start. Runoff is the portion of rainfall that flows over the surface rather than soaking in or evaporating. Peak flow is the maximum rate of discharge at a point during a storm, usually measured in cubic metres per second or cubic feet per second. Attenuation means reducing that peak, typically by storing water temporarily and releasing it more slowly. A combined sewer carries both sewage and stormwater in the same pipe; a separate system carries them in different pipes. A combined sewer overflow, or CSO, is a designed relief point where a combined sewer discharges excess flow, sewage included, to a river during heavy rain. A return period, such as the "100-year storm," is a statement of probability: a storm with a one per cent chance of being equalled or exceeded in any given year. It is not a promise that such a storm happens once a century, and much of the confusion in public debate about flooding begins with that misunderstanding. The reader does not need a background in engineering or mathematics. Where equations appear, they are explained in words and their behaviour matters more than their algebra. The aim is for a planner, a council member, an architect, a student or an interested citizen to finish the book able to read a drainage proposal critically: to ask which storms it is designed for, what it does to the peak and the volume, what happens when those storms are exceeded, and how anyone knows. Those four questions, applied consistently, are most of what separates a stormwater strategy that will hold up in the coming decades from one that will not. Chapter 1. The Paved Catchment Before a city existed on any given site, rain that fell there met a surface that was built to receive it. Leaves and branches intercepted the first few millimetres and let much of it evaporate back to the air. Leaf litter and organic soil acted like a sponge, holding water in their pores. Roots, burrows and decayed root channels opened routes for water to sink deep into the ground. Depressions in the land, small hollows and wetlands, held more. What reached the stream did so slowly, much of it as groundwater that seeped into the channel days or weeks after the rain had stopped. The stream was fed steadily, rising modestly in storms and keeping a reliable flow through dry spells. A city replaces that surface with materials chosen for exactly the opposite property. Roofs, roads and pavements are designed to be dry, durable and quick to shed water. The engineering virtue of a good road surface, that it does not let water in to weaken its foundation, is the hydrological vice of the city as a whole. Everything that follows in this book stems from that inversion. Where the water goes Hydrologists describe the fate of rainfall as a budget with a few main entries: interception and evaporation from surfaces, transpiration by plants, infiltration into the soil, and runoff over the surface. In a temperate forest, the proportions vary widely with soil and climate, but a commonly cited comparison in United States federal stream restoration guidance published in 1998 describes a natural ground cover in which roughly 40 per cent of rainfall returns to the air through evapotranspiration, about half infiltrates into shallow and deep soil, and only around 10 per cent becomes surface runoff. The same comparison describes a landscape with three-quarters or more impervious cover sending over half the rainfall off as runoff, with infiltration shrinking to a small fraction. The precise numbers depend on the site, and the comparison is schematic, but the direction and scale of the shift are not in doubt. Paving a catchment can multiply the volume of surface runoff several times over. Volume is only half of the change. The other half is speed. Water flowing over smooth asphalt or through a concrete gutter moves far faster than water working its way through grass and leaf litter. Once it enters a pipe, it can travel at one to three metres per second. A drop of rain that might have taken hours to reach a stream in a forest can reach the outfall of a storm drain in minutes. Engineers capture this with the idea of time of concentration: the time it takes for water from the hydraulically most distant part of a catchment to reach the outlet. Urbanisation shortens it dramatically, and the consequences of that shortening are central to everything in later chapters. Think of a catchment as a collection of thousands of small areas, each sending its runoff to the outlet with its own travel time. The flow at the outlet at any moment is the sum of contributions arriving at that moment. If travel times are long and widely spread, contributions from different parts of the catchment arrive at different times, and the peak is low and broad. If travel times are short and similar, contributions pile up together, and the peak is tall and sharp. Paving and piping both shorten and compress travel times. The result is a hydrograph, the curve of flow over time, with a much higher peak arriving much sooner after the rain. Field studies comparing urban and rural catchments have long found peak flows from heavily urbanised basins several times higher than from comparable undeveloped ones for the same storm, with the multiplier largest for the small, frequent storms that a natural landscape would have absorbed almost entirely. That last point is often missed. For an extreme storm, even a natural catchment becomes saturated and produces a lot of runoff, so the relative difference made by paving is smaller. For a modest storm, say ten or fifteen millimetres of rain, a forest might produce almost no runoff while a car park produces nearly the full amount. The biggest proportional change that urbanisation makes to hydrology is in the frequent storms. This is why the stream running through a growing suburb begins to flood several times a year when once it flooded rarely, and why its banks erode, its bed scours down, and its ecology collapses well before anyone notices a change in the size of the hundred-year flood. The urban stream syndrome Ecologists coined the phrase "urban stream syndrome" in the mid-2000s to describe the consistent pattern of decline in streams draining urban land. The symptoms are recognisable worldwide: flashier hydrographs with higher and more frequent peaks, elevated concentrations of nutrients and contaminants, altered channel shape, and a loss of sensitive species. Research associated with the concept, notably work by Christopher Walsh and colleagues in Melbourne, pointed to a specific mechanism that reshaped how the problem is understood. Ecological condition correlated less with the total amount of impervious surface in a catchment than with the amount that was directly connected to the stream by pipes. A roof that drains onto a lawn behaves very differently from a roof that drains through a downpipe straight into the storm sewer. The first has its runoff partly absorbed by soil; the second delivers every drop, rapidly, to the stream. This idea, called effective or connected imperviousness, matters for engineering because it identifies the lever. The problem is not paving as such. It is the efficient hydraulic connection between paved surfaces and receiving waters. The Melbourne research suggested that even quite low levels of connected impervious cover, a few per cent of the catchment, were enough to produce measurable ecological harm in small streams. If that is right, then no realistic amount of restraint in paving will protect urban streams. What can protect them is breaking the connection: intercepting runoff close to where it is generated and returning a portion of it to the soil and the air, as the natural landscape would have done. That is the purpose of much of the green infrastructure discussed in later chapters. The same reasoning explains the regulatory shift in many countries from controlling peak flow alone to controlling volume and frequency. Older rules required developers to build a detention pond that limited the peak discharge from a site to the pre-development peak for, say, the ten-year storm. This protected against the obvious flood, but it did nothing about the total volume, which was still much larger than before, and it often did nothing for the small storms, which passed straight through the pond's outlet. The stream continued to see far more frequent high flows than it would have naturally, and continued to erode. Newer standards, such as those requiring the retention on site of a defined depth of rainfall, typically the 85th or 95th percentile storm, aim to reproduce the natural water balance for most storms, not just the peak of a single design event. Combined and separate sewers Most of the world's older cities were drained before anyone thought to separate rainwater from sewage. In the nineteenth century, the great sanitary reforms in London, Paris, Hamburg, Chicago, New York and elsewhere built sewers that carried both in the same pipe. The logic was sound for its time. Rain flushed the sewers. Building one network was far cheaper than building two. And the outlet was a river that, before the age of treatment works, received the sewage whether it rained or not. When treatment arrived in the twentieth century, these combined systems acquired an awkward feature. Treatment works were sized to handle the sewage flow plus a modest allowance for rain, often a few times the average dry-weather flow. A sewer network that might carry fifty times the dry-weather flow during a storm could not send all of it to treatment. So the networks were equipped with overflow structures: weirs at which flow above a certain level spills into a river, carrying a dilute but real mixture of sewage and stormwater. These combined sewer overflows were deliberate pressure valves. They protected homes from having sewage back up into basements by sending it to the river instead. In the United States, the Environmental Protection Agency estimates that around 700 communities, home to tens of millions of people, are served by combined sewer systems, concentrated in the older cities of the Northeast and the Great Lakes region. In England, the sewers serving most of the country's towns and cities are combined in their older parts, and the public reporting of storm overflow spills since the late 2010s has made sewage pollution one of the most contested environmental issues in British politics. London's combined sewers, designed by Joseph Bazalgette after the Great Stink of 1858, had by the early 2000s been discharging tens of millions of cubic metres of combined sewage into the tidal Thames in a typical year. The scale of that problem, and the decision to build a tunnel beneath the river to fix it, is the subject of a later chapter. In separate systems, stormwater pipes carry only rain, and sewage goes to treatment in its own pipes. This avoids overflows of raw sewage, at least in principle, though in practice cross-connections and leakage often let rainwater into sanitary sewers and sewage into storm drains. Separate storm drains bring their own pollution problem, because runoff from streets carries oil, metals from brake pads and tyres, sediment, bacteria from animal waste, fertilisers and litter. Untreated, this all goes to the receiving water. For a long time stormwater was regarded as relatively clean. It is now understood as a major source of diffuse pollution, and much of the impetus behind green infrastructure in separately sewered cities is water quality rather than flood control. The distinction between combined and separate systems shapes the engineering in a fundamental way. In a combined system, the overriding goal is usually to reduce overflow volume and frequency, and every cubic metre of rain kept out of the sewer, or held back until treatment capacity is free, counts directly toward that goal. In a separate system, the goals are usually flood protection, stream protection and pollutant removal. The same bioswale can serve all of these, but the numbers that justify it, and the size it needs to be, differ. Two networks, one surface The final feature of urban drainage that needs establishing is the distinction between what engineers call the minor and major systems, a vocabulary that has been standard in North American and Australian practice for decades and appears under other names elsewhere. The minor system is the network of inlets, gutters and pipes that carries the runoff from frequent storms. It is typically designed for a storm with a return period of somewhere between two and ten years, sometimes thirty in critical locations. The major system is everything that carries the flow when the minor system is full: streets, overland flow paths, open channels, parks, and any other route the water finds. Every city has a major system whether or not anyone designed it. When the pipes fill, water runs on the surface along the path of steepest descent. If that path has been thought about, it might be a road with raised kerbs that acts as a channel and delivers the water safely to a river. If it has not, it might run through a row of houses, pond in an underpass, or pour into a metro station entrance. The deaths in basement apartments in New York during Ida and in the metro in Zhengzhou were failures of the major system: places where surface water accumulated in spaces where people were, with no route out. The minor and major systems interact at every inlet and manhole. When the pipe is under capacity, surface water drains into it. When the pipe is surcharged, meaning pressurised above its crown, water can be forced back out of manholes and gullies onto the street, so the pipe adds to the surface flood rather than relieving it. Any honest account of urban flood risk has to treat the two systems together, and much of the progress in the field over the past twenty years has come from models that do exactly that. What follows from this The picture that emerges from this account has several features that recur through the rest of the book. Urbanisation increases both the volume and the speed of runoff, and so produces higher and earlier peaks. The proportional change is largest for small, frequent storms, which matter most for stream health and water quality. The decisive factor is not paving in itself but the hydraulic connection of paved surfaces to receiving waters. Older cities have combined sewers that overflow sewage into rivers in wet weather, which makes volume control, not just peak control, an urgent public health goal. And every drainage network has two layers, a buried minor system for routine storms and a surface major system for extreme ones, whose interaction determines who gets flooded and how badly. Each of these observations points toward a different intervention. Restoring infiltration and evapotranspiration targets the volume and the frequent storms. Temporary storage targets the peak. Deep storage targets overflow volume in combined systems. Surface design targets the extreme storms that exceed everything else. The rest of the book follows those threads. But first it has to deal with the other half of the problem, the one that urbanisation did not create but has made far more dangerous: the rain itself is changing. Chapter 2. The Changing Rain Every drainage system ever built embodies a guess about the future. When an engineer sizes a pipe, she chooses a storm that the pipe will carry and accepts that larger storms will overwhelm it. That choice is expressed in the language of probability, as a return period or an annual exceedance probability, and the numbers behind it come from the historical record of rainfall at gauges nearby. For most of the twentieth century this was a sound procedure, because there was little reason to think that the statistics of rain would change over the life of the pipe. That assumption has now failed, and understanding why it failed, and what is replacing it, is essential to every design decision discussed in later chapters. How engineers describe rain Drainage design does not use the rain of any real storm. It uses a statistical abstraction called the intensity-duration-frequency relationship, usually abbreviated IDF. The idea is straightforward. From a long record at a rain gauge, extract the largest rainfall in each year for a range of durations: the wettest five minutes, the wettest hour, the wettest six hours, the wettest day. For each duration, fit a probability distribution to those annual maxima. The fitted distribution tells you, for example, that a one-hour rainfall of 45 millimetres has a 10 per cent chance of being equalled or exceeded in any year, which is what the "10-year one-hour rainfall" means. Repeat for every duration and a family of curves emerges: for each return period, intensity falls as duration increases, because the heaviest bursts of rain are short. The choice of distribution is not trivial. Classical practice often used the Gumbel distribution, a member of the family of extreme value distributions that has a fixed shape. Modern practice generally uses the generalised extreme value distribution, which adds a shape parameter that controls how heavy the upper tail is. That parameter matters enormously for rare events. Two distributions that fit thirty years of data equally well can disagree by a large margin about the 100-year or 500-year rainfall, because those estimates are extrapolations far beyond the observed record. National atlases address this by pooling data across many gauges in a region, on the reasoning that gauges in a climatically similar area share a common shape even if their magnitudes differ. NOAA Atlas 14, the standard reference across most of the United States since the mid-2000s, uses regional frequency analysis of this kind, as does the Flood Estimation Handbook used in the United Kingdom since 1999. An IDF curve is not a storm. To route water through a model of a drainage network, engineers need a hyetograph, a time series of rainfall intensity over the course of an event. Design practice converts IDF data into synthetic design storms. Some are simple blocks of uniform intensity. Others have a peak in the middle and tapering arms, built so that every sub-duration within the storm matches the IDF depth for the chosen return period. The Chicago design storm, published by Keifer and Chu in 1957, was an early version of this idea. The US Soil Conservation Service, later the Natural Resources Conservation Service, produced dimensionless 24-hour distributions labelled Type I, IA, II and III for different regions, which are still widely used. British practice developed its own profiles, with summer and winter variants. The choice of hyetograph is a hidden design decision with large consequences. A storm that concentrates its rain in a short central burst produces a high peak flow from a small, fast catchment, and is the critical case for sizing pipes. A storm that spreads the same depth over many hours produces a larger total volume over the period that matters for a big storage tank, and is the critical case for sizing storage. Engineers therefore test a range of durations to find the critical storm for each element. A small car park might be critical for a fifteen-minute storm; a large detention basin for a six-hour or twenty-four-hour one. When the rainfall statistics change, they may change differently for different durations, and that unevenness matters a great deal. What a return period does and does not mean The vocabulary of return periods is misunderstood so routinely, by the public and by decision-makers, that it is worth stating precisely. A "100-year storm" is a storm with a one per cent probability of being equalled or exceeded in any single year. It does not happen once every hundred years. Two can happen in consecutive years, or in the same week. Over a longer horizon the probabilities accumulate in ways that surprise people. If the annual probability is one per cent and years are independent, the chance of at least one such storm occurring in a 30-year period is about 26 per cent. Over 50 years it is about 39 per cent, and over 100 years about 63 per cent. A house protected to the 100-year standard has roughly a one-in-four chance of seeing that standard exceeded during a typical mortgage. For the storms that drive the design of pipe networks, the numbers are starker. A sewer designed for the 10-year storm has a 10 per cent annual chance of being exceeded, which makes exceedance a near certainty over the life of the pipe; the chance of at least one exceedance in 30 years is about 96 per cent. That is not a failure of design. It is the intended outcome. Pipes are designed for the frequent storm and the surface is expected to handle the rare one. The problem is that the second half of that sentence was often left unplanned. Because of this confusion, many agencies now prefer annual exceedance probability, or AEP, to return period. A "1% AEP" event says what it means. This book uses both terms, since both are in common use, but the reader should always translate a return period into an annual probability before reasoning about risk. Why warming intensifies short-duration rain The physical basis for expecting heavier rain in a warmer climate is the Clausius-Clapeyron relation, derived in the nineteenth century, which describes how the amount of water vapour air can hold at saturation increases with temperature. Near typical surface temperatures the increase is roughly 7 per cent per degree Celsius. If the strongest rainfall events are limited by the amount of moisture available to be rained out, their intensity should increase at about that rate as the atmosphere warms. Average rainfall is constrained by the global energy budget and increases more slowly, around 1 to 3 per cent per degree, which is why the extremes are expected to change faster than the mean. Observations broadly support this. Analyses of daily rainfall records across the globe find increases in the heaviest daily totals over most land regions with good data, and the Intergovernmental Panel on Climate Change, in its Sixth Assessment Report of 2021, concluded that the frequency and intensity of heavy precipitation events have increased since the 1950s over most land areas with sufficient observational data, and that human-induced climate change is likely the main driver. The same report presented estimates of how a heavy daily rainfall event that occurred once in ten years in a climate without human influence would change: at 1 degree of global warming, roughly the present level, such an event becomes about 1.3 times as frequent and about 7 per cent wetter; at 2 degrees, about 1.7 times as frequent and 14 per cent wetter; at 4 degrees, nearly 2.7 times as frequent and about 30 per cent wetter. For urban drainage, the more important finding concerns short durations. Small urban catchments respond to rainfall over minutes to a few hours, not days. In 2008, Geert Lenderink and Erik van Meijgaard published an analysis of a long record of hourly rainfall at De Bilt in the Netherlands showing that the most intense hourly rainfall increased with temperature at roughly twice the Clausius-Clapeyron rate over part of the temperature range. This "super-Clausius-Clapeyron" scaling has since been found in many, though not all, locations and is commonly attributed to the dynamics of convective storms: warmer, moister air releases more latent heat when it condenses, strengthening the updraft, which draws in more moisture and rains it out more efficiently. The scaling is debated, and its strength varies by region, season and storm type. But the weight of evidence, including from high-resolution convection-permitting climate models that can simulate individual thunderstorms, suggests that the short, intense bursts that overwhelm urban drainage are likely to intensify faster than daily totals. That matters because it is exactly the rainfall that urban drainage is most sensitive to. A sewer designed against a 30-minute intensity that has risen by 20 or 30 per cent is not slightly overloaded. Because flow in a full pipe increases with the square root of the head driving it, a modest increase in inflow can produce a large increase in surcharge and in the volume escaping to the surface. And because urban systems have little spare storage, the excess has nowhere to go. Stationarity is dead In 2008, a group of hydrologists led by P. C. D. Milly published a short article in Science with a provocative title: "Stationarity Is Dead: Whither Water Management?" Stationarity is the assumption that the statistical properties of a natural process, its mean, its variance and the probability of its extremes, do not change over time. Every IDF curve fitted to a historical record assumes it. The article argued that human influence on climate, along with changes in land use, had undermined that assumption to the point that it should no longer be the default basis for water resource planning. It did not offer a replacement. It said, in effect, that the profession needed one. The practical consequences surfaced in several ways. First, the design rainfall values in many national atlases were found to be out of date. In Illinois, for example, the State Water Survey updated the long-standing design rainfall statistics in March 2019 in a publication known as Bulletin 75, replacing the values of Bulletin 70, which had been in use since 1989. For Chicago, the 100-year 24-hour rainfall rose from 7.58 inches to 8.57 inches, an increase of about 13 per cent in thirty years of additional record, with knock-on effects for floodplain mapping and stormwater ordinances. Communities that adopted the new values found that infrastructure built to the old standard now offered a lower level of protection than intended. Second, agencies began to adopt explicit climate change uplifts. In England, the Environment Agency publishes allowances for peak rainfall intensity that drainage designers must apply to their design storms. Since a revision in May 2022 these allowances are given by river management catchment, with separate central and upper-end values for the 3.3 per cent and 1 per cent annual exceedance probability events and for epochs centred on the 2050s and 2070s. Upper-end allowances for the 1 per cent event in the 2070s are commonly around 40 per cent in many catchments, and higher in some. Danish practice has long applied climate factors to design rainfall, and Copenhagen's cloudburst planning after 2011 incorporated an allowance for future intensification into its design event. Other cities and countries have adopted similar uplift factors, often in the range of 10 to 40 per cent depending on the design life and the risk tolerance of the asset. Third, the national atlases themselves began to change. In the United States, NOAA began work on a replacement for Atlas 14, designated NOAA Atlas 15, with a key methodological shift: abandoning the stationarity assumption. The new atlas is planned in two volumes. The first is to provide current precipitation frequency estimates using nonstationary statistical methods that account for trends in the historical observations; the second is to provide projections of future precipitation frequency through the end of the century by applying adjustment factors derived from climate model output. A pilot covering Montana was released in preliminary form in September 2024. At the time of writing, NOAA's schedule calls for preliminary estimates for the contiguous United States no earlier than late 2026 and published estimates no earlier than late 2027, so most American designers will be working with Atlas 14 values plus local adjustments for some time yet. Designing without a fixed target If the rain is changing, what should an engineer design for? There is no single answer, but several principles have emerged that shape the rest of this book. The first is that the design life of the asset determines the climate horizon that matters. A permeable pavement with a 20- to 30-year surface life faces a different future from a deep tunnel with a design life of 120 years. Applying the same uplift to both makes little sense. Infrastructure that is expensive to modify later, such as tunnels, major trunk sewers and the levels of buildings and roads, deserves the most conservative assumptions. Infrastructure that can be adapted incrementally, such as street-level green infrastructure, can be designed for nearer-term conditions with a plan to add capacity as evidence accumulates. The second is that exceedance must be planned for regardless. Even a perfect forecast of future rainfall statistics would still imply that the design storm will be exceeded with some probability. The question is not whether the system will be overwhelmed, but where the water goes when it is, and whether that place is safe. A city that cannot answer that question for its 1 per cent or 0.5 per cent AEP storm has not finished its drainage design, however good its pipes are. The third is that uncertainty about future rainfall is large and is not going to be resolved quickly. Climate models disagree about local changes in short-duration extremes, observations are short and noisy, and natural variability can mask or exaggerate trends for decades. Designers should therefore prefer solutions that perform reasonably across a range of futures over solutions that are optimal for one assumed future and brittle in others. This argues for flexibility, modularity, and for layered systems in which the failure of one element degrades gracefully into the next. The final chapter of this book returns to these principles in detail. The fourth is that the short-duration intensification matters most for the parts of the system that respond quickly, the small catchments, inlets, gutters and local pipes, while changes in multi-hour or multi-day volume matter most for large storage. A deep tunnel sized for a long, soaking storm is relatively insensitive to whether the most intense 15 minutes become 30 per cent heavier. The surface inlets of a city street are acutely sensitive to it. An adaptation strategy needs to examine each component against the duration to which it is sensitive. With these principles in hand, the book can turn from the rain to the water on the ground, and to the central technical question: how storage changes the shape of the flood. Chapter 3. Shaping the Hydrograph The central claim of this book is that urban stormwater management is the craft of controlling when water arrives. Flood damage depends on the peak of the hydrograph, the maximum rate of flow, because that is what overtops pipes and channels. Overflow pollution and stream erosion depend largely on the volume and the frequency of high flows. Both can be changed by storage, because storage lets water arrive at one rate and leave at another. This chapter sets out the basic quantitative tools for describing that process. They are older than most of the infrastructure discussed in later chapters, and most of them can be understood without heavy mathematics. What matters is their logic, and the way that logic changes when storms grow larger than the ones a system was designed for. From rainfall to peak flow The oldest tool in the urban drainage engineer's kit is the rational method, which in its modern form dates to the work of the Irish engineer Thomas Mulvaney in the 1850s and was popularised for urban sewers by Emil Kuichling in Rochester, New York, in 1889. It states that the peak flow from a small catchment equals the product of a runoff coefficient, the rainfall intensity, and the catchment area. In metric units with intensity in millimetres per hour and area in hectares, the peak flow in cubic metres per second is that product divided by 360. A two-hectare car park with a runoff coefficient of 0.9, meaning that 90 per cent of the rain runs off, receiving rain at 100 millimetres per hour, generates a peak flow of about half a cubic metre per second. The method's insight lies in which intensity to use. The rational method says to take the rainfall intensity for a storm whose duration equals the time of concentration of the catchment. The reasoning is that the peak occurs when the whole catchment is contributing at once, which first happens when the storm has lasted long enough for water from the most distant point to arrive. Shorter storms are more intense but end before the whole area contributes; longer storms engage the whole area but at lower intensity. The storm whose duration matches the time of concentration produces the largest peak. This reveals in miniature why urbanisation is so damaging. Paving raises the runoff coefficient, often from 0.2 or 0.3 for lawn or woodland to 0.9 for roofs and roads. Piping shortens the time of concentration, which means the relevant intensity is read from a shorter duration on the IDF curve, and shorter durations have higher intensities. Both effects multiply the peak. A catchment whose coefficient rises from 0.3 to 0.8 and whose time of concentration falls from an hour to fifteen minutes can see its peak flow increase by a factor of five or more for the same return period. The rational method has serious limits. It gives only a peak, not a hydrograph, so it cannot be used to size storage. It assumes uniform rainfall across the catchment and a coefficient that does not change during the storm, when in reality soils wet up and absorb less as a storm progresses. Most guidance restricts it to catchments of a few tens or hundreds of hectares. For larger areas and for any problem involving storage, engineers need a way to produce the whole hydrograph. Losses and the shape of runoff Producing a hydrograph takes two steps. The first is to decide how much of the rain becomes runoff, which hydrologists call the loss model. The second is to decide how that runoff is distributed in time as it travels to the outlet, the routing or transformation. The most widely used loss model in North America, and influential elsewhere, is the curve number method developed by the US Soil Conservation Service in the 1950s. It assigns each combination of land cover and soil type a curve number between about 30 and 100, where higher numbers mean more runoff. Soils are classified into four hydrologic soil groups, from A, deep sands and gravels that infiltrate rapidly, to D, clays and shallow soils over rock that infiltrate very slowly. The method then computes runoff depth from rainfall depth through an empirical equation that includes an initial abstraction, the rain absorbed before any runoff begins, and a maximum retention that is approached as the storm continues. The equation's behaviour illustrates the arguments of the first chapter vividly. For a 3-inch storm, about 76 millimetres, falling on woodland in good condition on a sandy group B soil, with a curve number of 55, the method predicts about 5 millimetres of runoff, less than a tenth of the rain. The same storm on a residential area with quarter-acre lots on the same soil, curve number about 75, produces about 24 millimetres. On a commercial district, curve number around 92, it produces about 55 millimetres. On fully paved ground, with a curve number of 98, about 70 millimetres, more than nine-tenths of the rain. Development on the same soil has multiplied runoff volume by more than ten. On clay soils, where even the woodland runs off substantially, the multiplier is smaller, between two and three, which confirms the pattern seen in Chapter 1: urbanisation changes the hydrology of permeable landscapes most dramatically. Other loss models are used in other places. Continuous models represent soil moisture explicitly, using the Horton or Green-Ampt equations for infiltration rate, which fall as the soil wets. British practice developed its own percentage runoff relationships in the Wallingford Procedure of the early 1980s. The choice matters less than the recognition that pervious surfaces in a city contribute runoff that varies sharply with storm size and antecedent wetness, and that this nonlinearity is a large source of uncertainty in any model. Once runoff volume is known, it must be spread over time. The classic tool is the unit hydrograph, introduced by Leroy Sherman in 1932: the hydrograph of direct runoff produced at the outlet by one unit of runoff depth generated uniformly over the catchment in a given duration. If the catchment behaves linearly, the response to any storm can be built by scaling and adding unit hydrographs for each time step of rainfall excess. Synthetic unit hydrographs, such as the SCS dimensionless unit hydrograph, allow this to be done for catchments without streamflow records, using estimates of lag time from slope, length and land cover. In urban models, a different approach is more common: each subcatchment is treated as a nonlinear reservoir, a sloping plane that fills with rain and drains according to the equation for shallow overland flow. This is the method used by the United States Environmental Protection Agency's Storm Water Management Model, SWMM, the most widely used urban drainage model in the world, discussed in the next chapter. How storage attenuates a peak Now introduce storage. Suppose a hydrograph enters a detention basin, a pond or tank with a controlled outlet. Inflow and outflow need not be equal at every moment; the difference goes into or comes out of storage. This is simply conservation of mass: the rate of change of stored volume equals inflow minus outflow. The outflow, however, is not free to vary independently. It depends on the water level in the basin, because the outlet is a fixed structure, an orifice or a weir or a pipe, whose discharge increases as the water level above it rises. The water level in turn depends on how much is stored. So outflow depends on storage, and storage depends on the history of inflow minus outflow. The procedure for solving this, called level-pool or storage-indication routing, is usually attributed to the American engineer L. G. Puls in the 1920s and 1930s and is sometimes called the modified Puls method. It steps through time, computing at each step the new storage from the inflow and the previous outflow, then reading off the new outflow from the stage-storage-discharge relationship of the basin. The mathematics is not the point here. Three consequences of the procedure are. The first consequence is that the peak outflow always occurs where the outflow hydrograph crosses the falling limb of the inflow hydrograph. Before that point, inflow exceeds outflow and storage is rising, so the level and the outflow are still climbing. After it, outflow exceeds inflow, storage is falling, and the outflow declines. So the maximum outflow is lower than the maximum inflow, and it occurs later. That is attenuation and delay, the two effects that make storage valuable. The second consequence is that the amount of attenuation depends on the ratio of storage volume to the volume of the inflow hydrograph, and on how the outlet constrains discharge. A basin that is large relative to the flood, with a small outlet, produces heavy attenuation: the outflow is a low, long plateau. A basin that is small relative to the flood, or with a large outlet, barely changes the hydrograph. Note that a detention basin with an outlet does not reduce the total volume that leaves it, unless some water infiltrates or evaporates. It redistributes that volume in time. A retention facility, by contrast, holds some of the water permanently or until it infiltrates or is used, and so reduces the volume as well as the peak. The distinction between detention and retention recurs throughout this book. The third consequence, and the one least appreciated, is the behaviour of the system when the storm exceeds the design. Every storage element has a maximum level, beyond which an emergency spillway engages or the structure overtops. Below that level, the outlet throttles discharge and the basin absorbs the excess. Once the basin is full, it no longer absorbs anything; inflow passes straight through, over the spillway, at nearly its full rate. The attenuation does not decline gradually as storms grow. It collapses at the point of fullness. A worked example A simple calculation makes these effects concrete. Consider a detention tank with vertical walls and a plan area of 2,500 square metres, about the size of a third of a football pitch, draining through a 400-millimetre circular orifice at its base, with an overflow weir five metres long set at a depth of two metres. Into it flows a simplified triangular hydrograph, the kind a modest urban catchment might produce: flow rises linearly from zero to a peak of 2.0 cubic metres per second at thirty minutes, then falls linearly to zero at ninety minutes. The total inflow volume is 5,400 cubic metres. Routing this hydrograph through the tank, step by step, gives the following. The outflow peaks at about 0.40 cubic metres per second, one-fifth of the inflow peak, and it does so at around 78 minutes, nearly three-quarters of an hour after the inflow peak. The water in the tank reaches a maximum depth of about 1.63 metres, corresponding to about 4,100 cubic metres in storage. The overflow weir never engages. The tank has cut the peak by 80 per cent and delayed it by almost 50 minutes, which for a downstream sewer or stream is exactly the change needed: a smaller flow, arriving after the peaks from other tributaries have already passed. Now increase the storm. Suppose the inflow peak is 20 per cent higher, 2.4 cubic metres per second, with the same timing, a change in line with the climate uplifts discussed in the previous chapter. The tank's maximum depth rises to just under two metres, almost exactly at the weir crest, and the outflow peak rises to about 0.45 cubic metres per second. The tank still works; it has simply used its remaining freeboard. The outflow is only modestly higher, because the orifice discharge grows with the square root of the water depth, and that is a gentle function. Now increase the inflow peak to 3.0 cubic metres per second, 50 per cent above the original. The tank fills to its crest before the inflow peak has passed, the overflow weir engages, and the outflow peak jumps to about 1.3 cubic metres per second, more than three times the original design outflow. About 1,200 cubic metres pass over the weir. The attenuation has fallen from 80 per cent to about 55 per cent, and the remaining reduction comes only from the volume already captured before the tank filled. The example is idealised, but its lesson is general and is the core reason this book insists on layered systems. Storage elements behave very well up to their capacity and poorly beyond it. A system designed exactly for today's design storm, with little freeboard, has no margin for the intensification expected over its life. And the performance cliff is steep. The engineer cannot treat a 20 per cent increase in rainfall as producing a 20 per cent increase in downstream flow; depending on where the system sits relative to its capacity, it may produce a 10 per cent increase, or a 200 per cent increase. Timing across a network A single storage element is easy to reason about. A network of them is not, and here intuition fails frequently. Peak flow at a downstream point depends not only on the size of each tributary's peak but on their timing. If two tributaries peak simultaneously, their peaks add. If one peaks well before the other, the combined peak is smaller than the sum. Storage delays peaks, which is generally beneficial for the element's own outlet. But delaying a tributary's peak can, in the wrong place, shift it into coincidence with the peak from another tributary, increasing the combined flow downstream. This is a well-documented effect in catchment-scale flood studies. Detention ponds placed on tributaries near the bottom of a catchment, whose natural peaks would have passed before the main stem peak arrived, can delay them into the main peak and make downstream flooding worse. Conversely, storage placed in the upper parts of a catchment, whose peaks naturally arrive late, delays them further, past the main peak, and helps. Whether distributed storage helps or harms the catchment as a whole depends on where it is and on the storm. This insight has a practical corollary that many stormwater ordinances ignore. Requiring every new development to limit its own peak discharge to the pre-development rate, the most common regulatory approach in North America for decades, does not guarantee that peak flows downstream are unchanged. The volume released by each site is greater than before, and released over a longer period, so the flood recession is prolonged. If many sites release prolonged flows that overlap at a downstream point, the peak there can rise even though every individual site complies. Studies of catchments developed under such ordinances have found exactly this. The solution, which better regulations now adopt, is to control volume as well as peak, and to plan storage at the catchment scale rather than parcel by parcel. A further complication arises in pipe networks. Pipes themselves provide storage, sometimes a great deal of it, in the volume between the invert and the crown. A network of large trunk sewers may hold tens of thousands of cubic metres when full. That in-system storage attenuates peaks in the same way a tank does, but it is not controlled. Real-time control, the subject of a later chapter, is the practice of using gates and pumps to control it. What the tools tell us The quantitative tools introduced here support a few robust conclusions. Peak flow depends on rainfall intensity at the duration that matches the catchment's response time, and urbanisation increases peak flow by raising runoff volume and shortening response time simultaneously. Storage reduces and delays peaks, with an effectiveness governed by the ratio of storage volume to flood volume and by the outlet design. Detention changes timing but not volume; retention changes both. Storage performance collapses at the point of fullness, which makes the margin between the design storm and plausible future storms the most important number in any storage design. And at the catchment scale, the timing of attenuated peaks determines whether distributed storage helps or hurts. None of these conclusions can be acted upon for a real city using hand calculations. Real networks have thousands of pipes, hundreds of storage elements, surcharging and backwater, and surface flows that leave the pipes entirely. The next chapter turns to the hydraulic models that engineers use to represent all of this, and to the critical question of when their outputs can be trusted. Hashtags: #UrbanHydrologyAndGreenInfrastructure #UrbanHydrology #GreenInfrastructure #StormwaterManagement #ChangingClimates #UrbanFlooding #Hydrograph #PeakFlowAttenuation #Runoff #ImperviousSurfaces #ConnectedImperviousness #CombinedSewerOverflows #PermeablePavements #Bioswales #DetentionStorage #RetentionSystems #DeepTunnelStorage #SurfaceFloodRoutes #MajorMinorDrainageSystems #IntensityDurationFrequency #ClimateChangeUplifts #NonstationaryRainfall #RationalMethod #HydraulicModelling #FutureOfClimateResilientStormwater