This article examines how states today win influence less by threat and more by appeal. Building on Joseph Nye's idea of #soft_power, first set out in 2004 and revised in his later work, it asks a simple question: why do some countries get others to want what they want without paying or pushing? To answer this, the study reads soft power through three social theories that are rarely placed side by side. From Pierre Bourdieu it borrows #cultural_capital and #symbolic_power, which explain how taste, education, and prestige become quiet forms of advantage. From world-systems theory it borrows the #core_periphery model, which shows that attraction flows along the same lines as money and trade. From institutional theory it borrows #institutional_isomorphism, which explains why states copy each other's cultural programs until they all look alike. The article uses a qualitative, theory-led comparison of two well-documented cases: China's #Confucius_Institutes and South Korea's #Korean_Wave. The analysis finds that soft power is real but fragile, that it depends heavily on credibility, and that state control often weakens the very attraction it is meant to build. The study argues that #attraction works best when it is believed, not announced, and offers a framework for studying influence as a form of cultural currency. Keywords: soft power; cultural capital; world-systems theory; institutional isomorphism; public diplomacy; cultural diplomacy; global politics; nation branding 1. Introduction For most of recorded history, power in #global_politics was measured in armies, ships, and money. A strong state was one that could force a weaker state to do something, or pay it to comply. This view is not wrong, but it is incomplete. States also get their way through a third route, one that involves neither #military_coercion nor cash. They get their way because other people, and other governments, find them appealing. A student in Lagos who learns Korean to follow her favorite band, a young engineer in Nairobi who studies in Beijing on a scholarship, a film audience in Cairo that comes to admire Japanese design, each of these is a small unit of influence that was never bought and never forced. Added together, across millions of people, these small units shape what countries trust, buy, copy, and support. Joseph Nye named this resource #soft_power. He defined it as the ability to get the outcomes you want through attraction rather than payment or threat. The currency of this kind of power is not gold or guns. It is #culture, political #values, and the example a country sets through its own behavior. A state that is admired does not have to spend as much to get cooperation, because cooperation is freely given. Nye's later writing stresses that soft power has not faded even in an age of war and hard bargaining; governments that ignore it lose a tool that costs little and lasts long. The aim of this article is to take this familiar idea and give it a sharper theoretical edge. Soft power is often described in everyday language, as if attraction were simply a matter of good movies and friendly leaders. That description is true but shallow. It does not explain why attraction follows certain paths, why it can be inherited like wealth, or why governments around the world keep building the same kinds of cultural centers, scholarship schemes, and #nation_branding campaigns. To explain these patterns, this article reads soft power through three bodies of social theory. The first is the sociology of Pierre Bourdieu, whose ideas of #cultural_capital and #symbolic_power show how prestige and taste work like a hidden form of money. The second is world-systems theory, associated with Immanuel Wallerstein, which divides the world into a wealthy #core, a struggling #periphery, and an in-between #semi_periphery, and which suggests that cultural influence flows along the same channels as economic power. The third is the institutional theory of Paul DiMaggio and Walter Powell, whose concept of #institutional_isomorphism explains why organizations, and now states, tend to copy each other until their cultural programs look almost identical. The article then tests these ideas against two of the most studied soft power projects of the last two decades: China's #Confucius_Institutes and South Korea's #Hallyu, or #Korean_Wave. These cases were chosen because they are different in almost every way. One is led tightly by the state; the other grew first from private industry and fans. One has met growing suspicion in Western countries; the other has spread with little political resistance. Comparing them shows what attraction can and cannot do. The rest of the article is organized in the standard way. Section 2 sets out the theoretical framework. Section 3 describes the method. Section 4 presents the analysis of the two cases through the three theories. Section 5 reports the main findings. Section 6 concludes with the wider lesson: that #soft_power is best understood as a currency that loses value the moment people stop believing in it. 2. Background and Theoretical Framework 2.1 Soft power: the original idea and its recent life Nye built the concept of soft power on a simple split between three ways to move others: you can threaten them (sticks), you can pay them (carrots), or you can attract them (magnets). The first two are #hard_power. The third is soft. The sources of soft power, in his account, are a country's culture, its political values, and its foreign policy, but only when these are seen as honest and fair by outsiders. Importantly, soft power is not owned by the government alone. Much of it sits in society, in universities, studios, charities, and ordinary citizens, which means a state can never fully control it. In his more recent writing, Nye has answered his critics directly. Some argued that soft power is too vague to measure, or that it never really changes what states do. He responds that the concept was always meant to describe a resource, not a guaranteed result, and that the line between attraction and pressure can blur into what others have called #sharp_power, the use of cultural and information tools to manipulate rather than to charm. This distinction matters for the cases below, because a program that begins as attraction can curdle into something that foreign publics experience as interference. Two points from this literature guide the rest of the article. First, #credibility is the hidden engine of soft power. A message attracts only when the audience trusts the messenger, which is why propaganda usually fails. Second, soft power is relational: it exists in the eye of the receiver, not in the intention of the sender. A government can spend heavily and still produce no attraction at all if its actions contradict its words. 2.2 Bourdieu: attraction as cultural capital Pierre Bourdieu never wrote about international relations, yet his tools fit the subject well. He argued that power in society is not only economic. People also hold #cultural_capital, which includes education, manners, taste, language, and the credentials that signal belonging to a respected group. This capital can be converted into real advantage, much as money is, and it tends to be passed down and reproduced over time. Alongside it sits #symbolic_power, the quiet ability to make one's own standards look natural and universal, so that others accept them without noticing they are doing so. Recent scholarship has carried these ideas into the study of global governance and authority. Work on how business actors gain authority in world institutions, for example, shows that some bodies accumulate enough cultural and symbolic capital to set the terms that others must follow, framing their private interests as the public good (Calmettes & Pfister, 2025). The reach of Bourdieu's vocabulary across the social sciences has itself been mapped, confirming that concepts such as #field, #habitus, and symbolic capital remain among the most widely used tools for studying power (Schirone, 2023). Read this way, soft power is the projection of national cultural capital. When a country's universities are seen as the best, when its language is treated as the language of success, when its films define what is modern and desirable, that country holds symbolic power. It does not have to argue for its values; its values feel like common sense. This is why education and #higher_education are such strong soft power tools, and why scholarships are an investment in future loyalty. The student trained abroad carries a set of dispositions, a #habitus, that may quietly favor the host country for the rest of a career. 2.3 World-systems theory: attraction follows the money World-systems theory offers a colder, more structural view. It holds that the modern world is a single capitalist economy split into three zones. The #core contains rich, powerful states that set the rules and capture most of the value. The #periphery supplies cheap labor and raw materials and stays poor. The #semi_periphery sits in between, exploited by the core but exploiting the periphery in turn. Hegemony, in this model, is the position of the leading core power that shapes the system's institutions to suit itself. The theory was built to explain economics, and critics have long noted that it underplays #culture. Yet recent work has pushed it back into debates about power shifts, arguing that the old core-periphery hierarchy now runs through technology standards, finance, and rival blocs, and describing a phase of #interregnum in which a declining hegemon and rising challengers like China and Russia form a contested #semi_core (Demirel, 2024). This update matters for soft power because it locates attraction inside a structure. Cultural influence does not float free; it flows mostly downhill, from core to periphery, along the same channels as capital and trade. This framing explains an awkward fact. The countries with the most admired culture are usually also the richest and most powerful. American films, British universities, and French fashion are attractive partly on their merits, but also because they ride on economic dominance. When a rising power like China tries to build soft power, it is, in effect, trying to convert economic strength into cultural #legitimacy, to move from being a feared producer to an admired model. World-systems theory predicts this will be hard, because the existing core already defines what counts as good taste, and peripheral and semi-peripheral states may receive a newcomer's culture with suspicion rather than admiration. 2.4 Institutional isomorphism: why everyone copies everyone The third lens explains a pattern that is easy to see but rarely named. Around the world, governments now run almost identical soft power machinery: a cultural institute named after a famous figure, a scholarship program, a nation-branding agency, a glossy index ranking. Why this sameness? DiMaggio and Powell answered a version of this question for organizations decades ago, and they returned to it in a recent reassessment of their own theory (Powell & DiMaggio, 2023). Their concept of #institutional_isomorphism describes how organizations in the same field grow alike through three pressures. #Coercive_isomorphism comes from rules and powerful patrons that force conformity. #Mimetic_isomorphism comes from uncertainty: when no one is sure what works, everyone copies the leader. #Normative_isomorphism comes from shared training and professional norms that spread the same playbook. A concise restatement of the concept confirms its core claim, that organizations converge to gain #legitimacy rather than to improve performance (Sakib, 2022). Applied to states, this theory predicts exactly what we observe. A country that wants respect copies the soft power tools of admired powers, not because it has proof they work, but because using them signals that it is a serious, modern actor. The result is a global field of look-alike #cultural_diplomacy programs. This is a crucial insight, because it means much soft power spending is driven by the wish to look legitimate, not by any measured effect on foreign opinion. It also helps explain a recurring disappointment: programs copied for the sake of conformity often fail to attract, because attraction cannot be mass-produced. 2.5 Bringing the three together Each theory captures something the others miss. Bourdieu explains the content of attraction, the prestige and taste that make a culture desirable. World-systems theory explains the structure, the economic ladder along which attraction travels. Institutional isomorphism explains the form, the copied programs through which states try to manufacture appeal. Together they turn Nye's everyday idea into a testable framework: soft power is national cultural capital, projected along the channels of the world economy, through institutional forms that states copy from one another. The two cases below put this framework to work. 3. Method This study uses a qualitative, theory-led comparative case design. It does not run experiments or surveys of its own. Instead it builds an interpretive argument by reading well-documented cases through a defined set of concepts, a method common in #international_relations and political sociology when the goal is explanation rather than measurement. The evidence base is the recent peer-reviewed literature on soft power, #public_diplomacy, and #cultural_diplomacy, with priority given to studies published within the last five years so that the analysis reflects the current state of the field. From this literature, two cases were selected through purposive sampling: China's Confucius Institutes and South Korea's Korean Wave. The selection follows a "most different" logic. The two cases share an outcome of interest, a deliberate national effort to build attraction, but differ sharply on the key variable of state control. China's program is centrally directed; Korea's grew first from private industry and audiences and was only later embraced by the state. Comparing cases that differ on this dimension helps reveal what role state direction plays in whether soft power succeeds. The analysis proceeds in three steps. First, each case is described in plain terms, drawing on the recent literature. Second, each case is read through the three theoretical lenses set out above, asking what cultural capital it projects, where it sits in the #core_periphery structure, and how far its institutional form has been copied from others. Third, the two readings are compared to draw out general findings about how the currency of attraction gains and loses value. The study has clear limits, which honesty requires stating. It relies on secondary sources rather than new fieldwork, so it inherits their biases. Soft power outcomes are hard to measure, and any claim about cause and effect should be read with caution. Two cases cannot represent every form of #national_influence; programs in the Gulf, India, or the European Union would add other lessons. The aim, therefore, is not final proof but a clearer way of thinking, a framework that future empirical work can test. 4. Analysis 4.1 China's Confucius Institutes: state capital meets a credibility gap China's bid for #soft_power is one of the largest cultural projects of the century. At its center stood the Confucius Institutes, language and culture centers attached to foreign universities and named after China's most revered teacher. The logic was straightforward and, on paper, sound. Language is a gateway to cultural capital. Teach the world Mandarin, introduce it to Chinese art and thought, and a generation will grow up able and willing to engage with China, easing trade, study, and diplomacy. The program was tied closely to wider efforts such as the Belt and Road Initiative, where shared language smooths cooperation. Read through Bourdieu, the strategy makes sense. China was trying to build and export national cultural capital, to make its language and traditions a recognized form of prestige. Yet the program ran into a problem that Bourdieu himself would predict. #Symbolic_power cannot simply be claimed; it must be granted by others. And the source of the message shaped how it was received. Because the institutes were funded and overseen by the state, many foreign hosts came to see them not as neutral cultural centers but as arms of government policy. Recent scholarship on host-country reactions argues that the state-led design carried the seeds of its own difficulty: the more visibly official the program, the more suspicion it attracted, especially in Western democracies, with a wave of closures following in the United States and elsewhere (Li, 2024). The deeper damage shows up at the level of individual perception. A detailed study comparing Nye's society-based model of soft power with a more state-directed Chinese model found that, at the cognitive level, China's cultural programs were often perceived negatively and judged ineffective by the very audiences they targeted, a result the authors describe as a kind of #cognition_wall (Han & Cappelletti, 2025). In other words, the harder the state pushed, the less it attracted. This is the slide from soft power into #sharp_power that Nye warned about: when attraction is engineered too openly, audiences feel managed rather than charmed, and #credibility collapses. World-systems theory adds the structural reason this was always going to be uphill. China was, in these terms, a rising #semi_core power trying to convert economic weight into cultural legitimacy while the existing core still defined good taste and trustworthy speech (Demirel, 2024). Within established core societies, China's message met a hostile interpretive frame. The response has been a strategic shift. Rather than abandon the effort, Beijing has redirected much of its cultural and educational outreach toward the Global South, where scholarships, exchanges, and infrastructure ties find a more receptive audience and where China's example as a development model carries real appeal. Institutional theory explains the program's very design. The Confucius Institute was, in form, a copy of older national cultural bodies that admired powers had long operated. This is #mimetic_isomorphism at work: facing uncertainty about how to build influence, China adopted the recognized template of the cultural institute because it signaled membership in the club of serious global actors. A study of Chinese academic outreach in Morocco frames part of this dynamic as #policy_borrowing, the adoption of educational forms that travel between states (Yi & Shunu, 2025). Yet copying the form did not copy the effect. The template, designed for and trusted from established core states, did not automatically generate attraction when used by a newcomer. Even efforts framed as gentle "pragmatic enticement," offering tangible benefits like training and opportunity, show that the program worked best where it delivered concrete value rather than abstract image-making (Repnikova, 2022). 4.2 South Korea's Hallyu: attraction that the state did not invent The Korean Wave offers an almost mirror-image case. #Hallyu refers to the global spread of South Korean popular culture: music, television drama, film, food, and beauty products. Its most visible faces are global music acts and award-winning films, but its base is far wider, reaching into language learning, tourism, and consumer goods across every region. The crucial fact, easy to miss, is that Hallyu did not begin as a government plan. It grew first from private media companies chasing profit and from overseas audiences who simply liked the content. The state came later, recognizing a national asset and supporting it through policy, funding, and #cultural_diplomacy. This origin gave Hallyu the one thing China's program struggled to buy: credibility. Because the attraction came first from industry and fans rather than from official channels, foreign audiences did not experience it as propaganda. They experienced it as entertainment they had chosen. Through Bourdieu, Hallyu is a striking case of building new cultural capital almost from scratch. Korea was not a traditional cultural superpower. Yet within two decades it made its popular culture a recognized marker of what is modern, stylish, and desirable, especially among the young. Korean became a language worth learning not because a ministry urged it, but because fans wanted to understand the songs and shows they loved. This is symbolic power earned through appeal rather than asserted through authority, and it has lifted Korea's standing on global soft power rankings. The discourse around Hallyu also reveals how states try to capture cultural energy after the fact. A discourse analysis of Korean news and social media shows that the language of soft power became a way for the country to celebrate its rising global status, with conservative media in particular linking cultural success to national economic strength and using it to position Korea favorably against rivals (Yoon, 2023). Government and industry have since folded Hallyu into a deliberate strategy, integrating cultural exports with diplomacy, education, and trade promotion. The rise of domestic #digital_platforms has become a soft power apparatus in its own right, carrying Korean content worldwide and giving the state new ways to support its spread (Jin, 2024). World-systems theory frames the larger achievement. Korea moved up the global ladder, using cultural export as one way to consolidate a position above the periphery and to claim a respected place in the order. Crucially, Korean content traveled not only "down" to peripheral audiences but also "up," winning admiration inside the established core itself, in Europe and North America. This upward flow is unusual and is part of what makes the case important: it shows that the core-periphery structure shapes attraction but does not fully fix it. Institutional isomorphism appears here too, but in reverse order. China copied the institutional form and then sought the content; Korea generated the content and then built the institutional form to support it. As Korea's success became visible, other states began to copy Korea, studying the "K-model" of culture-led influence. The Korean Wave has itself become a template that others now imitate, an example of how a successful innovator reshapes the whole field and sets new #normative_isomorphism standards for what good cultural diplomacy looks like. 4.3 Reading the two cases together Placed side by side, the cases sharpen each other. Both states sought the same currency of #attraction. Both projected national cultural capital. Both operated within a global economic structure that channels influence. Both used, or were copied into, recognizable institutional forms. The decisive difference was the relationship between the state and the source of attraction. Where the state was the visible author of attraction, as with the Confucius Institutes, the attempt triggered suspicion and the message lost credibility, sliding toward the manipulation that audiences resist. Where attraction arose first in society and the state followed, as with Hallyu, credibility held and the influence proved durable, even reaching into the core. This contrast is the heart of the analysis and the basis for the findings below. 5. Findings The comparison yields five findings, each tying back to the theoretical framework. First, soft power is real but conditional. Both cases confirm that #attraction is a genuine resource that shapes study choices, consumer habits, language learning, and goodwill. But it is not automatic. It is conditional on credibility, and credibility depends on who is seen to be speaking. Attraction that is believed creates influence; attraction that is doubted creates resistance. This supports Nye's own emphasis on credibility as the engine of soft power and clarifies a frequent confusion in policy debates that treat cultural spending as if it reliably buys affection. Second, state control is a double-edged tool. The cases show that heavy, visible state direction can poison attraction. The Korean Wave succeeded in part because the state did not invent it; the Confucius Institutes struggled in core markets in part because the state was their obvious author (Han & Cappelletti, 2025; Li, 2024). The lesson is not that government should stay out of cultural diplomacy, but that its hand should be light and its claims honest. The moment a cultural program reads as a political instrument, it tips from soft power toward sharp power and loses the trust that made it work. Third, attraction flows within an economic structure but is not trapped by it. World-systems theory rightly predicts that cultural influence travels mostly from richer, more powerful zones outward, and that a rising #semi_core power faces a wall of suspicion when it tries to claim cultural legitimacy (Demirel, 2024). China's experience in core markets fits this prediction. Yet Hallyu's success inside the core shows that the structure constrains without fully determining. Genuine quality and authentic appeal can move culture "upward" against the usual current, which means structural position shapes the odds but does not settle the outcome. Fourth, much soft power spending is driven by imitation, not evidence. The global sameness of cultural institutes, scholarship schemes, and #nation_branding campaigns is well explained by institutional isomorphism (Powell & DiMaggio, 2023; Sakib, 2022). States copy admired models to look legitimate, often through #policy_borrowing, rather than because they have measured what works (Yi & Shunu, 2025). This finding carries a warning. A program adopted to fit in, rather than designed for a specific audience, is likely to underperform, because attraction is particular and cannot be mass-produced from a generic template. Fifth, the most durable soft power is the kind a state cannot fully own. This is the deepest finding and the one that unites the others. The sources of lasting attraction, a vibrant culture, trusted institutions, an admired example, sit largely in society, not in government. Bourdieu's account of cultural capital and symbolic power suggests why: prestige must be granted by others to be real, and it cannot be commanded into being. The states that gain the most are those that cultivate the conditions for attraction and then let it grow, rather than those that try to engineer and control it. 6. Conclusion The title of this article calls soft power a currency, and the comparison shows why the image fits. Like any currency, attraction has value only because people believe in it. A banknote is just paper; it works because a community trusts that it is worth something. Soft power works the same way. A country's culture, values, and example are worth influence only as long as foreign publics find them credible and appealing. The instant that belief fades, the currency is devalued, no matter how much was spent to print it. This reframing turns Nye's well-known idea into a richer analytical tool. Bourdieu shows that soft power is national cultural capital, prestige that can be converted into real advantage but that must be granted, never simply claimed. World-systems theory shows that this capital tends to flow along the ladder of the world economy, giving established core powers a head start and making the rise of a challenger's culture an uphill task, though not an impossible one. Institutional isomorphism shows that states, anxious to appear legitimate, copy one another's cultural machinery until the tools look identical, which is why so much spending produces so little distinctive appeal. The two cases turn these abstractions into a clear practical lesson. South Korea's #Korean_Wave thrived because its attraction was authentic and largely beyond state control, which protected its credibility even inside the wealthy core. China's #Confucius_Institutes met resistance in those same markets because their state authorship invited suspicion, pushing the program toward the manipulation that audiences reject and prompting a strategic turn toward the Global South. The difference was not the size of the budget. It was the trust of the audience. The wider implication for #global_politics is both hopeful and sobering. It is hopeful because it means that influence need not rest on #military_coercion; a society that is genuinely admired holds a kind of power that is cheaper and more humane than force. It is sobering because that power cannot be faked, bought wholesale, or fully controlled. The states that will lead in the coming decades by attraction rather than by fear are likely to be those that invest in being worth admiring, and that have the confidence to let their culture and values speak for themselves. The currency of attraction, in the end, is earned, not minted. #soft_power #cultural_diplomacy #public_diplomacy #global_politics #cultural_capital #symbolic_power #world_systems_theory #institutional_isomorphism #Nye #Bourdieu #attraction #national_influence #nation_branding #Korean_Wave #Confucius_Institutes #TheCurrencyOfAttraction #SoftPowerInGlobalPolitics #SoftPowerStudies #InternationalRelations #PoliticalSociology #CulturalInfluence #GlobalAffairs #ForeignPolicy #SoftPowerResearch #STULIB References Calmettes, A., & Pfister, D. J. (2025). Constructing business authority in global governance: A Bourdieusian account of multi-level meaning fixation. Review of International Studies. https://doi.org/10.1017/S0260210525100922 Demirel, E. (2024). Anarchy in the world-system: The emergence of the semi-core and the phase of interregnum. Journal of World-Systems Research, 30(1), 373–397. https://doi.org/10.5195/jwsr.2024.1209 Durrani, S. (2023). What happens when a country bleeds soft power? Conceptualising 'Negative Watch': Towards an epistemology for negative and adversarial place branding. Place Branding and Public Diplomacy, 19(4), 530–544. https://doi.org/10.1057/s41254-023-00302-9 Gallarotti, G. M. (2022). Esteem and influence: Soft power in international politics. Journal of Political Power, 15(3), 383–396. https://doi.org/10.1080/2158379X.2022.2135303 Han, Y., & Cappelletti, A. (2025). Perceptions of China's soft power: Exploring the cognition wall through the case of the Confucius Institute. Humanities and Social Sciences Communications, 12, Article 1090. https://doi.org/10.1057/s41599-025-05359-w Jin, D. Y. (2024). The rise of digital platforms as a soft power apparatus in the New Korean Wave era. Global Media and China, 9(2), 153–168. https://doi.org/10.1177/20570473241234204 Li, S.-y. (2024). Problematising China's soft power: Setbacks of Confucius Institutes and the decentralisation of its international language and culture promotion strategy. The Pacific Review, 38(3), 397–421. https://doi.org/10.1080/09512748.2024.2387614 Nye, J. S. (2021). Soft power: The evolution of a concept. Journal of Political Power, 14(1), 196–208. https://doi.org/10.1080/2158379X.2021.1879572 Nye, J. S. (2023). Soft power. In Soft power and great-power competition (China and Globalization series, pp. 1–8). Springer. https://doi.org/10.1007/978-981-99-0714-4_1 Powell, W. W., & DiMaggio, P. J. (2023). The iron cage redux: Looking back and forward. Organization Theory, 4(4). https://doi.org/10.1177/26317877231221550 Repnikova, M. (2022). Rethinking China's soft power: "Pragmatic enticement" of Confucius Institutes in Ethiopia. The China Quarterly, 250, 440–463. https://doi.org/10.1017/S0305741022000340 Sakib, N. H. (2022). Institutional isomorphism. In A. Farazmand (Ed.), Global encyclopedia of public administration, public policy, and governance. Springer. https://doi.org/10.1007/978-3-030-66252-3_3932 Schirone, M. (2023). Field, capital, and habitus: The impact of Pierre Bourdieu on bibliometrics. Quantitative Science Studies, 4(1), 186–208. https://doi.org/10.1162/qss_a_00232 Yi, L., & Shunu, M. (2025). Academic diplomacy and policy borrowing: Media content analysis of Chinese soft power in Morocco. Humanities and Social Sciences Communications, 12, Article 399. https://doi.org/10.1057/s41599-025-03567-4 Yoon, K. (2023). De/constructing the soft power discourse in Hallyu. Communication Research and Practice, 9(3), 341–357. https://doi.org/10.1080/22041451.2023.2237811

Chronic diseases such as diabetes, hypertension, heart failure, and chronic kidney disease cause most deaths and absorb most health spending around the world. Many of the people who carry the heaviest burden also have the least access to regular care. This article asks a direct question: can #generative_AI chatbots and #remote_patient_monitoring tools narrow these gaps, or do they quietly widen them? Using the mixed-methods study by Andreadis and colleagues (2024) as a starting point, it reviews recent evidence on conversational agents and home-based monitoring for long-term conditions. The work follows an integrative review design. Sources published mainly between 2020 and 2025 were gathered, screened, and grouped into themes covering #diagnostic_detection, daily self-management, cost, and fairness. The themes were then read through three social theories: Bourdieu's ideas about #cultural_capital, the #world_systems perspective on core and periphery, and #institutional_isomorphism from organizational sociology. The analysis shows real promise. Chatbots can triage symptoms with accuracy close to, and sometimes above, older symptom checkers, while monitoring tools lower blood pressure and improve glucose control when paired with clinical follow-up. Yet the same study that praises chatbot accuracy also finds that advice shifts in tone and framing depending on a patient's age, sex, and race, which is a sign that old inequalities can be rebuilt inside new systems. The article concludes that these tools bridge gaps only when access, language, trust, and governance are treated as central design problems rather than afterthoughts. Keywords: generative artificial intelligence; conversational agents; remote patient monitoring; chronic disease management; diagnostic detection; digital health equity; large language models 1. Introduction Long-term illness is now the main story of global health. According to widely cited estimates, roughly one in four adults lives with at least one chronic condition, and these conditions drive a large share of hospital visits, disability, and early death (Zhang & Song, 2024). The standard model of care, built around scheduled clinic appointments, was never designed for diseases that need attention every day. A person with #diabetes makes dozens of small health decisions between visits. A person with heart failure can slide from stable to critical in the gap between two appointments. The result is a structural mismatch: continuous diseases met with episodic care. Two technologies have entered this gap with unusual speed. The first is the #generative_AI chatbot, powered by large language models that can hold a natural conversation, answer health questions, and suggest what a symptom might mean. The public release of conversational models in late 2022 pushed this from a research idea into everyday use, and surveys now report that a large minority of people already ask these tools about their health (Yau et al., 2024). The second is #remote_patient_monitoring, a setup where connected blood pressure cuffs, glucometers, scales, and pulse oximeters send readings from a person's home to their care team (Smith et al., 2024). One technology talks; the other watches. Together they promise something the clinic alone cannot offer: a presence in daily life. The phrase "bridging gaps" is doing a lot of work in this conversation, and it deserves a careful look. A gap in #access is not the same as a gap in #quality, and neither is the same as a gap in #trust. A tool can close one while leaving the others open, or even make them worse. This is why the study by Andreadis and colleagues (2024) is such a useful anchor. They compared a leading chatbot against a traditional online #symptom_checker across many simulated patients who differed only by age, sex, and race or ethnicity. Their work captures both sides of the promise at once: strong diagnostic performance sitting next to advice that subtly changes depending on who is asking. This article has three aims. First, to describe what current evidence actually says about chatbots and monitoring tools in #chronic_disease_management and diagnostic detection, in plain terms, without hype. Second, to explain why the same tool can help one group and disadvantage another, using social theory rather than only technical language. Third, to offer a grounded view of what "bridging gaps" would have to mean in practice for the claim to be honest. The contribution here is mostly conceptual. Rather than testing a new device, the article connects scattered clinical findings to a sociological reading of how technology spreads through health systems. Most reviews stay inside the language of accuracy and effect sizes. Fewer ask why a hospital in one country adopts the same #digital_health platform as a hospital on another continent, or why a well-designed tool can still leave a #marginalized patient worse served than a wealthy one. Those questions need theory, and the three frameworks chosen here are well suited to answering them. The timing matters. The current generation of conversational models is the first to be good enough that ordinary people use them by choice rather than as a novelty, and the first to be cheap enough that health systems can deploy them at scale. Decisions made in this short window, about how the tools are tested, who they are designed for, and which standards they must meet, will shape care for years, because health systems rarely unwind a technology once it sits inside their daily routines. Studying these tools now, while the choices are still open, is more useful than studying them later, once the choices have hardened into habit. That is the spirit in which this review is offered: not as a verdict on whether the technology is good or bad, but as an attempt to name the conditions under which it helps the people who need help most. The rest of the article is organized in the standard way. The next section sets out the background and the theoretical framework. After that comes the method, then the analysis of the evidence, then the consolidated findings, and finally the conclusion. 2. Background and Theoretical Framework 2.1 The technical background A large language model is a system trained on enormous amounts of text so that it can predict and produce fluent language. When wrapped in a chat interface, it becomes a #conversational_agent that can explain a diagnosis, answer follow-up questions, or walk someone through what to do about chest pain at two in the morning. Recent work shows that fine-tuned models can support the diagnosis of dozens of common chronic conditions from described symptoms, with reported accuracy on validation data that is, on paper, very high (Zhang & Song, 2024). Other studies pair these models with classical machine learning to classify and predict chronic disease risk, treating the chatbot as a friendly front door to a more complex engine underneath (Abbas et al., 2024). #Remote_patient_monitoring works on a different principle. Instead of generating language, it generates a steady stream of #physiological_data. A connected cuff records blood pressure each morning; a glucometer logs sugar levels; a scale flags sudden weight gain that can signal fluid building up in the body. The care team sees trends rather than single snapshots, which lets them act before a small problem becomes an emergency (Smith et al., 2024). A 2024 systematic review of monitoring programs found benefits across safety, treatment #adherence, quality of life, and cost, though the size of the benefit depended heavily on the condition and on how well the program was run (Tan et al., 2024). The two technologies are starting to merge. A chatbot can sit on top of a monitoring stream, reading a person's numbers and turning them into advice they can act on, such as a reminder, an explanation, or a nudge to call the clinic (Khetan & Suvarna, 2024). This is where "bridging gaps" sounds most convincing, because it offers both data and dialogue at the same time. 2.2 Why technical promise is not the whole story If the only question were accuracy, the conversation could end with the clinical data. But health technology does not float free of society. It is bought by institutions, paid for by insurers, used by people with very different resources, and built by companies based mostly in a handful of wealthy countries. To understand whether these tools bridge or widen gaps, the analysis needs to step outside the lab. Three theories help. 2.3 Bourdieu: capital, habitus, and symbolic power Pierre Bourdieu argued that people carry different forms of #capital: economic capital (money), social capital (useful relationships), and #cultural_capital (the knowledge, language, and confidence that let a person move smoothly through institutions) (Bourdieu, 1986). He also described #habitus, the deep set of habits and expectations shaped by where a person grows up, and #symbolic_power, the quiet authority that makes some ways of speaking seem more legitimate than others (Bourdieu, 1984). This framework explains a stubborn pattern in #digital_health. A chatbot rewards the user who knows how to ask a precise question, who reads at a comfortable level, and who feels entitled to push back when an answer seems off. That user is drawing on cultural capital. Someone with less of it may accept a vague answer, miss a warning sign, or never start the conversation at all. The Andreadis study makes this concrete: when the same medical question was asked by simulated patients of different backgrounds, the tool's answers shifted in tone and framing, even where the underlying diagnosis stayed similar (Andreadis et al., 2024). In Bourdieu's terms, the system was exercising symbolic power, treating some patients as more deserving of clear, complete, respectful explanation than others. A tool that does this does not simply transmit medical facts; it reproduces a #social_hierarchy while appearing neutral. This is why critics warn that models trained on human text tend to absorb the biases inside that text (Dankwa-Mullan, 2024). 2.4 World-systems theory: core, periphery, and unequal flows Immanuel Wallerstein described the global economy as a single system split into a wealthy #core, a poorer #periphery, and a semi-periphery in between, with resources and value flowing toward the core (Wallerstein, 2004). Applied to health technology, the pattern is hard to miss. The largest language models and the dominant monitoring platforms are designed, owned, and tuned in a small set of high-income countries. Their training data, clinical assumptions, and even their default language reflect that origin. When such tools are exported to the periphery, they can extend care to places short on specialists, which is a genuine good. But they also create dependence. A clinic in a lower-income region may rely on a platform it cannot inspect, adjust, or afford to replace, while the data it produces flows back toward the core (Kyriazakos et al., 2024). The #digital_divide is not only about whether a person has a smartphone; it is about who sets the standards that everyone else must follow. World-systems theory reframes "bridging gaps" as a question of power, not just connectivity. A bridge built and owned by one side of a river is still a bridge, but it is not a neutral one. 2.5 Institutional isomorphism: why everyone adopts the same tool DiMaggio and Powell asked why organizations in the same field come to look so alike, and named three pressures (DiMaggio & Powell, 1983). #Coercive_isomorphism comes from rules and funders: when a payer rewards monitoring programs, hospitals adopt them. #Mimetic_isomorphism comes from uncertainty: when leaders are unsure what works, they copy respected peers, so one prestigious system's AI rollout becomes the template for many. #Normative_isomorphism comes from professions: as training, conferences, and journals treat a tool as standard, using it becomes a mark of being a serious, modern provider. These pressures explain the speed of adoption far better than evidence of benefit does. Surveys show many health organizations experimenting with #generative_AI even while admitting the clinical proof is still thin (Yau et al., 2024). Isomorphism warns that this herd movement can spread an immature tool widely before its risks are understood, and can lock in a single vendor's design choices, including any biases baked into them. It also explains why fixing problems is hard once a tool becomes the field's default. Together, the three theories give the article its spine: Bourdieu explains unequal experience at the level of the individual, world-systems explains it between regions, and isomorphism explains how a single design, good or flawed, comes to dominate a whole field. 3. Method This article uses an #integrative_review design. An integrative review is a structured way of gathering, comparing, and combining different kinds of evidence, including clinical trials, systematic reviews, and conceptual work, in order to build a broader understanding of a topic. It suits a question that is partly empirical (do these tools work?) and partly social (for whom, and at what cost?). The search focused on peer-reviewed literature, with priority given to work published between 2020 and 2025 so that the clinical picture would reflect the current generation of #large_language_models and connected devices. Records were identified through major health and multidisciplinary databases using combinations of terms covering #generative_AI, conversational agents, chatbots, remote patient monitoring, telemonitoring, #chronic_disease, diagnostic accuracy, and #health_equity. The study by Andreadis and colleagues (2024) was treated as the anchor reference, and its themes guided the wider search. Sources were included if they reported on the use of chatbots or remote monitoring for long-term conditions, examined diagnostic detection by conversational tools, or analyzed access, bias, or fairness in #digital_health. Purely technical papers with no health application were set aside, as were opinion pieces without supporting evidence, except where they were needed to define a theory. Classic theoretical works by Bourdieu, Wallerstein, and DiMaggio and Powell were included deliberately despite their age, because foundational theory is judged by its explanatory power rather than its publication date, and no recent restatement replaces the originals. Analysis happened in two passes. The first was a #thematic_analysis: each source was read, and its main claims were tagged and then grouped into recurring themes, namely diagnostic detection, daily self-management, cost and system effects, and equity. The second pass was a #theoretical_synthesis, in which each theme was read against the three frameworks set out above to ask not only what the evidence showed but why the pattern might exist and whom it served. Two limits should be stated plainly. First, this is a narrative synthesis, not a meta-analysis, so it does not pool numbers into a single statistic; it describes the shape of the evidence rather than measuring it precisely. Second, the field moves quickly, and a model studied last year may already be outdated, so the conclusions are about patterns and directions rather than fixed performance figures. 4. Analysis 4.1 Diagnostic detection The clearest signal in the evidence is that #conversational_agents have become surprisingly capable at the first step of care, which is figuring out what a symptom might mean and how urgently it should be acted on. The Andreadis study found that a leading chatbot matched or outperformed a traditional online symptom checker on diagnostic suggestions across a wide range of common complaints (Andreadis et al., 2024). Purpose-built systems go further: a fine-tuned model designed for chronic conditions reported very high accuracy in identifying two dozen common diseases from described symptoms, and most users in its test found it convenient for everyday health questions (Zhang & Song, 2024). Hybrid approaches that combine a chatbot with machine learning have shown similar strength in classifying and predicting chronic disease risk (Abbas et al., 2024). These numbers should be read with care. High accuracy on a clean validation set is not the same as safe performance with a frightened patient using imperfect language at midnight. A study of chatbot answers to emergency questions found responses that were usually helpful but occasionally carried #dangerous_information, the kind of confident error that a worried person might act on (Yau et al., 2024). So the analysis points to a tool that is genuinely strong at #diagnostic_detection in controlled conditions, and genuinely risky when treated as a replacement for a clinician. 4.2 Daily self-management Where these tools may matter most is in the long, quiet middle of a chronic illness, far from any dramatic diagnosis. #Remote_patient_monitoring shows consistent benefit here. A large retrospective study of monitoring for high blood pressure found a clinically meaningful drop in average pressure among patients whose readings were tracked from home and reviewed by their team (Smith et al., 2024). A broad 2024 review across many programs reported gains in adherence, quality of life, and cost, strongest where monitoring was paired with real clinical follow-up rather than left to run on its own (Tan et al., 2024). Reviews of software-based monitoring tools point the same way, with the best results coming from programs that combine data with structured human support (Kyriazakos et al., 2024). Chatbots add a layer of dialogue on top of this data. They can explain a reading, send a reminder, answer a question that a patient would feel embarrassed to ask a person, and offer support at any hour (Khetan & Suvarna, 2024). The combination, data plus conversation, addresses the structural mismatch named in the introduction. It puts something between the appointments. The pattern in the evidence is that neither technology alone is as useful as the two working together with a clinician kept in the loop. 4.3 Cost and system effects The system-level picture is mixed but mostly hopeful. Monitoring programs can lower costs by catching deterioration early and preventing expensive emergencies, and a recent account of monitoring across the United States describes rapid growth driven partly by payment changes that reward it (Paul et al., 2025). This is #coercive_isomorphism in plain sight: when funders pay for a practice, the practice spreads. Chatbots promise to ease pressure on overstretched staff by handling routine questions, though the evidence that they actually reduce workload rather than shifting it remains thin (Yau et al., 2024). The honest reading is that these tools can save money and time under the right conditions, and can also add hidden costs in oversight, training, and error correction when rolled out carelessly. It is worth being concrete about where the hidden costs sit. A monitoring program that sends a daily stream of readings to a clinic also creates a daily stream of decisions: someone has to look at the data, judge which readings matter, and act on them. Without enough staff and a clear protocol, the data piles up unread, and the program delivers a false sense of safety rather than real protection. A chatbot raises a different version of the same problem. Each confident wrong answer it gives has to be caught and corrected somewhere downstream, often by the very clinicians the tool was meant to relieve. The benefit, then, is not a property of the device but of the system around it. The same monitoring kit that lowers costs in a well-staffed clinic can raise them in an understaffed one, which is one more reason that blanket claims about savings should be read with suspicion. 4.4 Equity, the thread through everything The most important theme cuts across all the others. Whether a tool bridges a gap or widens it depends on who can reach it and how it treats them once they arrive. Reviews of #digital_health_equity find that access to monitoring tools, apps, and wearables is uneven, with #underserved_communities consistently less likely to be enrolled (Adepoju et al., 2024). #Digital_health_literacy is a major barrier, especially for older adults in marginalized groups who may distrust the health system for good historical reasons (Owusu et al., 2024). And the Andreadis finding sits at the center of this thread: even when a chatbot reaches a patient, its advice can shift in tone and completeness based on demographic cues, which means the tool can deliver a quietly worse service to the people who already get worse care (Andreadis et al., 2024). Scoping work on equity frameworks argues that fairness has to be designed into every stage, not bolted on at the end (Kim & Backonja, 2025; Richardson et al., 2022). 5. Findings Reading the analysis through the three theories produces a set of findings that are more layered than a simple "these tools work" or "these tools fail." Finding 1: The tools are real bridges, but partial ones. On the technical evidence, both #generative_AI chatbots and #remote_patient_monitoring narrow specific gaps. Chatbots close a knowledge-and-access gap by offering credible #diagnostic_detection and round-the-clock answers (Andreadis et al., 2024; Zhang & Song, 2024). Monitoring closes a continuity gap by replacing occasional snapshots with steady data that prevents crises (Smith et al., 2024; Tan et al., 2024). The bridge is real. It is also incomplete, because it carries some travelers more comfortably than others. Finding 2: Bias is the price of fluency. A model that sounds human has learned from human text, and human text is uneven. The Andreadis result, where advice varied by patient demographics, is best understood through Bourdieu (Andreadis et al., 2024). The chatbot acts as a holder of #symbolic_power, granting fuller, clearer explanation to patients whose profile matches the confident, well-resourced user the data was richest in. Patients with less #cultural_capital receive thinner service and are less equipped to notice or challenge it. The danger is precisely that the bias is polite and invisible, wrapped in friendly language, which makes it harder to name than open rudeness would be (Dankwa-Mullan, 2024). Finding 3: The bridge mostly runs one way. Read through #world_systems thinking, the global pattern is a flow of tools outward from a wealthy core and a flow of data and dependence back toward it (Wallerstein, 2004). Tools built in one context carry that context's assumptions about language, diet, body, and disease, and exporting them without adaptation can misfire in the periphery even while appearing generous (Kyriazakos et al., 2024). Bridging a gap in #access can deepen a gap in #autonomy if the receiving system cannot inspect, adjust, or own the tool it now depends on. Finding 4: Adoption is racing ahead of evidence. Through the lens of #institutional_isomorphism, the speed of uptake looks less like a careful response to proof and more like organizations copying each other and responding to funders (DiMaggio & Powell, 1983). #Coercive pressure from payers, #mimetic copying of prestigious early adopters, and #normative pressure from the professions are pushing #digital_health forward faster than the clinical record justifies (Yau et al., 2024; Paul et al., 2025). The risk is that an immature or biased design becomes the field's standard before its flaws are understood, and then becomes hard to remove precisely because everyone uses it. Finding 5: Equity is a design choice, not a happy accident. Across every theme, the difference between bridging and widening came down to deliberate choices: who gets enrolled, in what language, with how much human support, and under whose governance (Adepoju et al., 2024; Owusu et al., 2024; Kim & Backonja, 2025). The tools that helped #marginalized patients did so because someone designed for them on purpose. Left to default settings, the technology drifted toward serving the already-served. Taken together, the findings reframe the article's title. #Generative_AI and monitoring tools can bridge gaps in chronic disease care, but "bridging" is not an automatic property of the technology. It is an outcome that has to be built, paid for, and defended against the pull of bias, dependence, and herd adoption. 6. Conclusion The evidence supports a careful kind of optimism. #Generative_AI chatbots can detect and explain disease at a level that would have seemed unlikely only a few years ago, and #remote_patient_monitoring can keep a clinical eye on people during the long stretches between visits, lowering blood pressure, steadying glucose, and catching trouble early (Andreadis et al., 2024; Tan et al., 2024; Smith et al., 2024). For the structural mismatch at the heart of #chronic_disease_management, continuous illness met by episodic care, these tools are a genuine answer. The same evidence refuses to let that optimism run unchecked. The study that anchors this article shows that the most fluent, accessible tool can still hand out unequal advice depending on who is asking (Andreadis et al., 2024). Read through Bourdieu, that is #symbolic_power reproducing #social_hierarchy in a new medium. Read through #world_systems theory, the global spread of these tools risks turning access into dependence. Read through #institutional_isomorphism, the rush to adopt them threatens to lock in flaws before anyone has measured them. None of these problems is a reason to reject the technology. Each is a reason to govern it. So the practical message is specific. To bridge gaps rather than rebuild them, designers and health systems should test tools for bias across the actual range of patients they serve, not just the easiest ones, and should publish what they find. They should pair every chatbot and monitoring stream with real human support, especially for patients with low #digital_health_literacy, instead of using technology to remove people from the loop. They should build #fairness into the design from the first day, treat #data_governance and ownership as matters of justice and not just compliance, and resist adopting a tool merely because respected peers have. Equity, in short, has to be the specification, not the marketing. Future work should follow patients over time, in ordinary clinics rather than ideal test conditions, and should pay particular attention to the people these systems are most likely to fail. The promise of bridging gaps with #generative_AI is real. Whether it is kept depends less on the models and more on the choices made by the humans who deploy them. Hashtags #GenerativeAI #Chatbots #RemotePatientMonitoring #ChronicDiseaseManagement #DiagnosticDetection #DigitalHealthEquity #LargeLanguageModels #HealthInformatics #AIinHealthcare #Telemonitoring #HealthDisparities #Bourdieu #InstitutionalIsomorphism #WorldSystemsTheory #DigitalDivide References Abbas, S., Iftikhar, M., Shah, M. M., & Khan, S. J. (2024). ChatGPT-assisted machine learning for chronic disease classification and prediction: A developmental and validation study. Cureus, 16(12), e75851. https://doi.org/10.7759/cureus.75851 Adepoju, O., Dang, P., & Nguyen, H. (2024). Equity in digital health: Assessing access and utilization of remote patient monitoring, medical apps, and wearables in underserved communities. Inquiry, 61, 1–8. https://doi.org/10.1177/00469580241271137 Andreadis, K., Newman, D. R., Twan, C., Shunk, A., Mann, D. M., & Stevens, E. R. (2024). Mixed methods assessment of the influence of demographics on medical advice of ChatGPT. Journal of the American Medical Informatics Association, 31(9), 2002–2009. https://doi.org/10.1093/jamia/ocae086 Bourdieu, P. (1984). Distinction: A social critique of the judgement of taste. Harvard University Press. Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241–258). Greenwood Press. Dankwa-Mullan, I. (2024). Health equity and ethical considerations in using artificial intelligence in public health and medicine. Preventing Chronic Disease, 21, 240245. https://doi.org/10.5888/pcd21.240245 DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101 Khetan, S. P., & Suvarna, S. S. (2024). Empowering patients through AI: The role of ChatGPT in daily monitoring of blood pressure and blood glucose levels. Global Cardiology Science and Practice, 2024(4), e202456. https://doi.org/10.21542/gcsp.2024.56 Kim, K., & Backonja, U. (2025). Digital health equity frameworks and key concepts: A scoping review. Journal of the American Medical Informatics Association, 32(5), 932–944. https://doi.org/10.1093/jamia/ocaf017 Kyriazakos, S., Pnevmatikakis, A., Kostopoulou, K., Ferrière, L., Thibaut, K., Giacobini, E., Pastorino, R., Gorini, M., & Fenici, P. (2024). Benchmarking the clinical outcomes of Healthentia SaMD in chronic disease management: A systematic literature review comparison. Frontiers in Public Health, 12, 1488687. https://doi.org/10.3389/fpubh.2024.1488687 Owusu, B., Gbaba, S., Juste, J., Bivins, B., & Baptiste, D. (2024). Digital health literacy and Black older adults: Bridging the digital divide in managing chronic conditions. Nursing Open, 11(12), e70123. https://doi.org/10.1002/nop2.70123 Paul, M. M., Khera, N., Elugunti, P. R., Ruff, K. C., Hommos, M. S., Thomas, L. F., Nagaraja, V., Garrett, A. L., Pantoja-Smith, M., Delafield, N. L., Lizaola-Mayo, B. C., Kresin, M. M., Seetharam, M., Nagarakanti, S. R., & Kaur, M. (2025). The state of remote patient monitoring for chronic disease management in the United States. Journal of Medical Internet Research, 27, e70422. https://doi.org/10.2196/70422 Richardson, S., Lawrence, K., & Schoenthaler, A. (2022). A framework for digital health equity. npj Digital Medicine, 5(1), 119. https://doi.org/10.1038/s41746-022-00663-0 Smith, W., Colbert, B. M., Namouz, T., Caven, D., Ewing, J. A., & Albano, A. W. (2024). Remote patient monitoring is associated with improved outcomes in hypertension: A large, retrospective, cohort analysis. Healthcare, 12(16), 1583. https://doi.org/10.3390/healthcare12161583 Tan, S. Y., Sumner, J., Wang, Y., & Yip, A. W. (2024). A systematic review of the impacts of remote patient monitoring (RPM) interventions on safety, adherence, quality-of-life and cost-related outcomes. npj Digital Medicine, 7(1), 192. https://doi.org/10.1038/s41746-024-01182-w Wallerstein, I. (2004). World-systems analysis: An introduction. Duke University Press. Yau, J. Y., Saadat, S., Hsu, E., Murphy, L. S., Roh, J. S., Suchard, J., Tapia, A., Wiechmann, W., & Langdorf, M. I. (2024). Accuracy of prospective assessments of 4 large language model chatbot responses to patient questions about emergency care: Experimental comparative study. Journal of Medical Internet Research, 26, e60291. https://doi.org/10.2196/60291 Zhang, S., & Song, J. (2024). A chatbot based question and answer system for the auxiliary diagnosis of chronic diseases based on large language model. Scientific Reports, 14, 17118. https://doi.org/10.1038/s41598-024-67429-4 Related tags: #BridgingGapsWithAI · #GenAI_in_ChronicCare · #AI_DiagnosticDetection · #RPM_for_ChronicDisease · #ConversationalAI_Health · #EquitableDigitalHealth · #AIChatbotsInMedicine · #DigitalHealthGap · #ChronicCareInnovation

Public hospitals across high-, middle- and low-income systems face the same hard problem: rising demand, tight budgets, and a clinical workforce that is stretched thin and expensive to expand. This article reviews the empirical evidence on two responses that are increasingly applied together, namely industrial #lean_management borrowed from manufacturing and #predictive_analytics drawn from data science. The aim is to map how these methods reduce #operational_waste and improve #clinical_staffing in government-run hospitals, and to read that evidence through three sociological lenses: Bourdieu's theory of #field and #capital, neo-institutional theory's account of #institutional_isomorphism, and Wallerstein's #world_systems_theory. Following the framing offered by Gherghina et al. (2024), the study uses an integrative review design, synthesising peer-reviewed work published mainly between 2021 and 2025 on lean adoption, kaizen, #value_stream_mapping, and machine-learning approaches to #demand_forecasting and #nurse_workload. The analysis finds consistent gains in #waiting_times, #length_of_stay, and bed throughput where lean is paired with daily management and frontline ownership, and growing evidence that forecasting models can match or exceed traditional staffing judgement when fed routine clinical data. It also finds that results are uneven, often fragile, and shaped less by technique than by professional power, imitation between organisations, and the global structure of the health labour market. The article argues that #waste_reduction in public hospitals is as much a social and political project as a technical one, and that ignoring this is the most common reason improvement programmes fade. Practical implications for managers, clinicians, and policymakers are discussed. Keywords: lean healthcare; public hospitals; predictive analytics; clinical staffing; operational waste; institutional theory 1. Introduction Public hospitals carry a double burden. They are expected to treat everyone who arrives, including the uninsured and the chronically ill, while operating under budgets set by ministries and treasuries rather than by markets. The combination of an ageing population, a heavier load of chronic disease, and patient expectations that keep climbing has put #public_hospitals under pressure that few of them were designed to absorb (Marsilio et al., 2022). When money is short and demand is not, two questions dominate management meetings. The first is how to stop wasting the resources already in the building. The second is how to get the right number of clinical staff to the right place at the right time. Industrial #lean_management offers an answer to the first question. Lean began on the factory floor of Toyota, where the goal was to remove anything that did not add value for the customer, including waiting, overproduction, unnecessary movement, and defects that have to be redone (Wang et al., 2025). From the early 1990s, hospitals began borrowing this thinking, treating the patient journey as a production process that could be mapped, measured, and smoothed. Tools such as #value_stream_mapping, #kaizen events, standard work, and visual management moved from car plants into emergency departments, pharmacies, and operating theatres (de Barros et al., 2021). The promise was simple and attractive: cut the #operational_waste that sits between a patient and their care, and you free up capacity without spending more. #Predictive_analytics speaks to the second question. Staffing a hospital is hard because demand is genuinely uncertain. Admissions rise with winter illness, with heatwaves, with road traffic, and with shifts in local population. Traditional rostering tends to react to yesterday's crisis rather than anticipate tomorrow's. #Machine_learning and other forecasting methods change the timing of the decision by using historical and real-time data to estimate how many patients will arrive, how sick they will be, and therefore how much nursing time a ward will need (Meredith et al., 2025; McMahon et al., 2025). When these forecasts are accurate, managers can plan base staffing weeks ahead and adjust surge staffing in near real time, reducing both dangerous understaffing and costly overstaffing. The interesting development of recent years is that these two tools are converging. Lean tells a hospital where its #patient_flow breaks down; analytics tells it when and by how much demand will move. Used together, they form a feedback loop in which #standardisation makes processes predictable enough to model, and models make staffing decisions precise enough to protect the gains from lean. This article, following the line of inquiry set out by Gherghina et al. (2024), examines the empirical evidence on that convergence in the specific setting of government-run hospitals. A purely technical reading of this evidence, however, misses most of the story. Hospitals are not factories. They are professional organisations in which doctors and nurses hold authority that no management chart can override, and they sit inside national and global systems that shape what counts as legitimate practice and where skilled staff can be found. For that reason this article does not stop at measuring whether lean and analytics work. It asks why they take hold in some places and stall in others, and it uses three sociological frameworks to answer. Bourdieu's concepts of #field, #habitus and #capital explain the clash between managerial and clinical authority. Neo-institutional theory explains why so many hospitals adopt the same methods regardless of fit. #World_systems_theory explains why the clinical staffing problem cannot be solved inside any single hospital, because the workforce itself flows along global lines of inequality. The article proceeds as follows. Section 2 sets out the theoretical framework. Section 3 describes the integrative review method. Section 4 analyses the evidence on lean and on analytics in turn. Section 5 presents the synthesised findings. Section 6 concludes with implications for practice and policy. 2. Background and Theoretical Framework 2.1 Lean as an industrial logic in a clinical setting Lean is best understood as a logic rather than a fixed toolkit. Its core claim is that most of the time and money lost in any system is spent on activities that the end user would never pay for if they could see them. In a hospital, that waste shows up as patients waiting for a bed, samples sitting in trays, staff walking long distances to find supplies, duplicated paperwork, and errors that send a case back to the start (Trakulsunti and Trakoonsanti, 2021). Lean methods make this waste visible and then attack it through #continuous_improvement, usually in small, repeated cycles led by the people who do the work. Two features of lean matter for the argument here. First, it was built for routine, high-volume, repeatable processes, which describes some hospital work, such as outpatient clinics and elective surgery, far better than it describes emergency or intensive care (Régis et al., 2019). Second, lean assumes that frontline workers will own the changes. Where management imposes lean from the top as a cost-cutting drive, the evidence shows it tends to be resented and to fade once attention moves on (Ivcovici et al., 2025). This tension between #standardisation and professional autonomy is exactly where the sociological lenses become useful. 2.2 Predictive analytics as the timing layer If lean reshapes the process, analytics reshapes the timing of decisions about that process. Forecasting #nurse_workload and patient arrivals turns staffing from a reactive task into a planned one. Recent work has shown that models built from routine data already collected in hospitals, such as ward occupancy, patient acuity, and nursing activity records, can predict workload with useful accuracy (Meredith et al., 2025; McMahon et al., 2025). In emergency settings, machine-learning models trained on millions of past visits can forecast arrivals and even rival the admission judgements of experienced triage nurses, and combining model output with nurse judgement can perform better than either alone (Nover et al., 2025; Porto and Fogliatto, 2024). The promise of analytics for #resource_allocation is real, but it depends on the same organisational conditions that lean does. A forecast only helps if the staffing system is flexible enough to act on it, and if clinicians trust it enough to follow it. A model that tells a manager to add a nurse is useless if there is no nurse to add, which returns the discussion to the global labour market addressed below. 2.3 Bourdieu: field, capital and the politics of improvement Pierre Bourdieu offers a way to understand why improvement programmes succeed or fail that has nothing to do with the quality of the tools. For Bourdieu, a hospital is a #field, a structured social space in which actors compete for position using different forms of #capital (Collyer et al., 2015). Doctors hold a particular kind of #cultural_capital grounded in clinical expertise and professional standing, a resource that managers cannot simply outrank. Nurses, administrators, and analysts hold other forms of capital with different and often lower exchange value inside the medical field. Lean and analytics enter this field as outsiders carrying managerial capital. When a manager presents a #value_stream_mapping exercise, they are not only proposing a process change; they are making a claim about whose knowledge counts. The concept of #habitus, the deeply held dispositions that shape how professionals see their work, explains why clinicians may experience standard work as an insult to their judgement rather than as a help (Ivcovici et al., 2025). Programmes that translate lean into terms that respect clinical capital, and that let clinicians lead the change, convert this resistance into ownership. Programmes that treat the hospital as a factory and clinicians as interchangeable workers run straight into the logic of the field and lose. This is the single most consistent lesson from the qualitative literature. 2.4 Institutional isomorphism: why everyone adopts the same thing Neo-institutional theory explains a puzzle that pure efficiency reasoning cannot. If lean worked everywhere, we would expect hospitals to adopt it only when it fits their problems. Instead, adoption often spreads in waves that look more like fashion than like fit. DiMaggio and Powell's classic account of #institutional_isomorphism describes three pressures that push organisations in the same field to resemble one another, independent of whether the change improves performance (DiMaggio and Powell, 1983). #Coercive_isomorphism comes from regulators and funders. When a health ministry ties accreditation or money to the use of quality-improvement methods, hospitals adopt lean to stay funded and legitimate (Lum et al., 2023). #Mimetic_isomorphism comes from uncertainty. When managers do not know how to fix a complex problem, they copy whichever flagship hospital is admired this year, importing its lean programme as a signal of seriousness (Marsilio et al., 2022). #Normative_isomorphism comes from professionalisation. As consultants, university programmes, and professional bodies spread a shared idea of what a modern hospital does, lean and analytics become part of the expected repertoire of any credible institution. The value of this lens is that it warns against reading adoption as evidence of usefulness. A hospital may run kaizen events because its neighbours do, while gaining little, and may even adopt the visible rituals of lean, such as boards and slogans, without the underlying discipline. This gap between adopted form and actual practice, often called decoupling, is a recurring theme in the findings. 2.5 World-systems theory: staffing as a global problem The third lens widens the frame from the single hospital to the planet. Wallerstein's #world_systems_theory divides the global economy into a wealthy core, a dependent periphery, and a semi-periphery in between, with resources and labour flowing from periphery to core (Wallerstein, 2004). Health labour follows this pattern. Nurses and doctors trained at public expense in lower-income countries migrate to wealthier systems that pay more, draining the very hospitals that can least afford the loss (Giwa and colleagues, 2025). This matters directly for the topic of the article. No amount of predictive analytics inside a peripheral hospital can forecast its way out of a structural shortage of nurses, if those nurses keep leaving for the core. Recent analysis of the Gulf states as a new "hybrid core" pulling health workers out of sub-Saharan Africa shows how #workforce_migration creates compound extraction pressures on poorer systems, which now lose staff to both the traditional Global North and to rising regional powers (Giwa and colleagues, 2025). Optimising #clinical_staffing, in other words, has a ceiling set far above the hospital level. Lean and analytics can make the best of the staff a hospital has, but they cannot manufacture staff that the world system keeps pulling elsewhere. Taken together, these three lenses convert a technical question into a social one. The rest of the article uses them to read the empirical evidence. 3. Method 3.1 Design This study uses an integrative review design, which is suited to synthesising evidence from different methods, including quantitative outcome studies, qualitative case studies, and review articles, around a focused question (de Barros et al., 2021). The guiding question was: what does the empirical literature show about how industrial #lean_management and #predictive_analytics reduce #operational_waste and optimise #clinical_staffing in public hospitals, and what social and institutional conditions shape those outcomes? The framing followed the agenda set out by Gherghina et al. (2024). 3.2 Selection of sources Sources were identified from major academic databases and reference chaining, prioritising peer-reviewed journal articles and systematic reviews. Three inclusion criteria were applied. First, relevance to lean, kaizen, process improvement, or predictive and machine-learning approaches to staffing and patient flow in hospital settings. Second, a focus on public, government-funded, or mixed public systems wherever possible, with private-sector studies included only for comparison. Third, recency, with priority given to work published between 2021 and 2025, while a small number of foundational theoretical texts were retained regardless of date because Bourdieu, Wallerstein, and DiMaggio and Powell are the conceptual anchors of the analysis. 3.3 Analytical approach Each source was read twice. The first reading extracted the empirical claims: what was done, in what setting, and what changed, with attention to measurable outcomes such as #waiting_times, #length_of_stay, throughput, cost, staffing accuracy, and #patient_safety. The second reading coded the social and institutional dimensions, asking who led the change, how clinicians responded, what external pressures drove adoption, and whether reported gains were sustained. These two layers were then mapped against the theoretical framework so that technical results and social explanations could be read side by side. 3.4 Limitations of the method An integrative review depends on published work, and publication favours success stories over failures, which almost certainly inflates the apparent #healthcare_efficiency benefits of lean and analytics. Many studies are single-site and short-term, measuring outcomes soon after an intervention when enthusiasm is high. Comparability is weak because hospitals define waste, acuity, and staffing differently. The review is also limited to sources in accessible languages, which skews toward certain regions. These constraints mean the findings should be read as a structured synthesis of the available evidence rather than as a pooled estimate of effect. 4. Analysis 4.1 What lean removes, and what it leaves behind The strongest and most repeated finding across the lean literature is that the method reduces time-based waste in routine, high-volume processes. Studies report shorter outpatient waiting times, reduced inpatient waiting, faster turnaround in pharmacy and procurement, and lower length of stay when lean tools are applied to a well-defined pathway (Trakulsunti and Trakoonsanti, 2021; Wang et al., 2025). A recent systematic review synthesising lean across efficiency, quality, cost, and satisfaction concluded that the method generally improves all four, with the clearest effects on operational efficiency and time-based measures (Wang et al., 2025). Work on specific clinical pathways, such as acute stroke care, similarly shows that lean redesign can compress the time from arrival to treatment, which in time-critical conditions translates directly into better outcomes (Cegłowska et al., 2024). These gains are real but conditional. The same literature shows that lean's effects depend heavily on how it is implemented. The Italy and United States benchmarking study found that hospitals reporting a stronger daily management system, meaning lean embedded in everyday routines rather than run as occasional events, reported more improvement attributable to lean (Marsilio et al., 2022). Studies of kaizen in Italian public hospitals found that initiatives leaning on the social side of lean, such as team engagement and shared problem-solving, produced different and often better outcomes than those leaning only on the technical tools. The implication is consistent: the toolkit alone does little, and the human and organisational scaffolding around it does most of the work. This is where Bourdieu's field becomes visible in the data. The qualitative study of an Australian public hospital showed hybrid clinical managers, people who are both doctors and managers, doing intense identity work to reconcile lean with their professional standing, and responding very differently depending on their contract and their voice in the process (Ivcovici et al., 2025). Where clinicians had real participation, lean was absorbed into professional practice. Where it was imposed, it was treated as a managerial intrusion. The technique was the same; the result depended on the politics of cultural capital. 4.2 Sustainability and the problem of fading A second analytical theme is that lean is far easier to start than to sustain. The Malaysian study protocol noted that although dozens of public hospitals had officially adopted lean, very little was known about whether the changes lasted, and identified sustainability as the central unanswered question (Lum et al., 2023). This is a quiet but damning point. Much of the literature measures outcomes during or just after a project, capturing a peak that may not survive once leaders move on, budgets tighten, or the original champions leave. Institutional theory predicts exactly this pattern. Where adoption is driven by #coercive_isomorphism, such as a ministry mandate, or by #mimetic_isomorphism, such as copying an admired hospital, the organisation has an incentive to display the visible signs of lean without committing to the harder daily discipline (DiMaggio and Powell, 1983). The result is decoupling: improvement boards on the wall, slogans in the corridor, and little real change in #patient_flow. Reading the sustainability gap through this lens suggests that fading is not a failure of the tool but a predictable outcome of adopting it for legitimacy rather than for use. 4.3 What predictive analytics adds The analytics literature is younger but moving fast. Several recent studies demonstrate that #nurse_workload can be predicted from routine hospital data with accuracy good enough to inform staffing. Models built on nursing activity documentation and on validated workload tools have produced credible next-shift and next-day workload estimates across multiple wards (McMahon et al., 2025; Meredith et al., 2025). In emergency care, machine-learning models trained on very large historical datasets forecast arrivals and admissions well, and studies combining model output with the judgement of triage nurses report better performance than either source on its own (Nover et al., 2025; Porto and Fogliatto, 2024). The operational logic is attractive. Forecasts made weeks ahead can set base staffing, while real-time predictions can guide surge decisions, allowing managers to match supply to demand rather than guess (Meredith et al., 2025). Studies that frame this as a prediction-driven staffing system report that even small shortfalls against recommended levels measurably increase patient waiting times, which makes the case for acting on accurate forecasts rather than on intuition. Yet analytics inherits every social constraint that limits lean, and adds a few of its own. A forecast is only useful if the staffing system can flex to meet it, which assumes a pool of available staff and contracts that allow rapid adjustment. It must also be trusted by the clinicians it instructs, which returns to Bourdieu: a model carries managerial and technical capital that may be resisted in a field where clinical judgement is the dominant currency. And it raises questions of fairness and transparency, since opaque models that shape who works when can be experienced as a loss of professional control. The most successful approaches in the literature are those that position the model as a support for clinical judgement rather than a replacement for it, which mirrors the lesson from lean about ownership. 4.4 Where lean and analytics meet The convergence the introduction described appears in the evidence as a reinforcing loop. Lean #standardisation makes processes regular enough to be modelled, because a pathway that varies wildly cannot be forecast. Analytics, in turn, protects lean gains by ensuring that the staffing needed to run the improved process is actually present at the right moment. Hospitals that pair the two report that #demand_forecasting tells them where surge capacity is needed, while lean tells them how to deploy it without creating new bottlenecks. The combination is more than the sum of its parts, but it also concentrates the demands on management capability, data quality, and clinical trust, which means the hospitals best placed to gain are often those that already function well. 5. Findings Synthesising the analysis, the review yields six findings. Finding 1: Lean reliably reduces time-based waste in routine processes, but the size and durability of the gain depend on implementation, not on the tool. Across settings, lean shortens waiting times and length of stay and improves throughput where it is applied to repeatable, high-volume pathways and embedded in daily routines rather than run as one-off events (Marsilio et al., 2022; Wang et al., 2025). The strongest predictor of success is the presence of a daily management system and genuine frontline ownership. Finding 2: Clinician engagement is the decisive variable, and it is best understood as a question of professional power. The qualitative evidence shows that the same lean methods succeed when clinicians lead and fail when they are imposed (Ivcovici et al., 2025). Bourdieu's account of #field and #cultural_capital explains why: improvement programmes that respect clinical authority are absorbed, while those that override it are resisted. #Waste_reduction is therefore as much a negotiation as a calculation. Finding 3: Much lean adoption is driven by legitimacy rather than by fit, which predicts both wide spread and frequent fading. The pattern of adoption matches #institutional_isomorphism better than rational choice. Coercive, mimetic, and normative pressures push hospitals to adopt lean regardless of need, and the resulting decoupling, where the visible signs of lean appear without the underlying discipline, explains why many programmes do not last (DiMaggio and Powell, 1983; Lum et al., 2023). Finding 4: Predictive analytics can match or exceed traditional staffing judgement, especially when combined with clinical expertise. Models built from routine data predict #nurse_workload and patient arrivals accurately enough to guide both base and surge staffing, and hybrid approaches that blend model output with clinical judgement outperform either alone (McMahon et al., 2025; Meredith et al., 2025; Nover et al., 2025). The value of a forecast, however, is bounded by the flexibility of the staffing system and the trust of the staff. Finding 5: Lean and analytics are complementary, forming a loop in which standardisation enables forecasting and forecasting protects standardisation. The combination offers more than either method alone, but it raises the bar for management capability, data quality, and clinical trust, which means the hospitals most able to benefit are frequently those already performing well. This risks widening the gap between strong and weak institutions. Finding 6: The clinical staffing problem has a structural ceiling that no hospital-level method can lift. #World_systems_theory shows that staff shortages in many public hospitals are produced by global #workforce_migration from periphery to core, including to new regional cores such as the Gulf states (Giwa and colleagues, 2025). Analytics can allocate scarce staff more wisely, but it cannot replace staff that the world system continually draws away. Optimising clinical staffing inside the hospital is necessary but not sufficient; the binding constraint sits at the level of national and global policy. Across these findings runs a single thread. The measurable, technical benefits of lean and analytics are genuine, but they are produced, limited, and sometimes erased by social and institutional forces, including the politics of professional power, the pressure to imitate, and the global structure of the health labour market. Treating #healthcare_efficiency as a purely operational matter is the most reliable way to lose the gains it produces. 6. Conclusion Public hospitals will keep facing demand that outruns their budgets, and they are right to look to proven methods for relief. The evidence reviewed here supports a qualified optimism. Industrial #lean_management genuinely reduces #operational_waste in routine care, and #predictive_analytics genuinely improves the precision of #clinical_staffing. Used together, they create a virtuous loop in which standardised processes can be forecast and forecast demand can be staffed. For managers and clinicians under pressure, these are real tools that can free real capacity. The argument of this article, however, is that technique is the smaller part of the story. Whether lean and analytics deliver depends on conditions that the tools themselves do not supply. Bourdieu shows that improvement runs through a #field shaped by professional #capital, so programmes that respect and empower clinicians succeed where those that override them fail. Neo-institutional theory shows that hospitals often adopt these methods to look legitimate rather than to solve a problem, which explains both their rapid spread and their tendency to fade once the spotlight moves. #World_systems_theory shows that the staffing crisis many public hospitals face is manufactured far above the ward, by a global market that pulls trained staff from poorer systems to richer ones. Three practical implications follow. First, leaders should treat lean and analytics as social interventions, investing as much in clinician ownership, daily routines, and trust as in the tools themselves, because the social scaffolding is what makes the technical gains last. Second, funders and regulators should be wary of mandating methods in ways that reward the appearance of #continuous_improvement over its substance, since coercive adoption produces decoupling rather than change. Third, policymakers should recognise that no hospital can analyse its way out of a structural workforce shortage; durable improvement in #clinical_staffing requires national training investment, ethical recruitment, and international agreements that slow the extraction of staff from systems that cannot spare them. Future research should follow lean and analytics programmes over years rather than months, report failures as readily as successes, and study public hospitals in peripheral and semi-peripheral systems on their own terms rather than as deficient versions of core-country models. Above all, it should keep the technical and the social in the same frame. The hospitals that reduce #waste_reduction to a spreadsheet exercise will keep rediscovering, at considerable cost, that a hospital is a community of professionals embedded in a wider world, and that lasting #healthcare_efficiency is built with that community, not imposed upon it. Hashtags #Lean_Healthcare #Public_Hospitals #Predictive_Analytics #Clinical_Staffing #Operational_Waste #Lean_Management #Healthcare_Efficiency #Hospital_Management #Kaizen #Value_Stream_Mapping #Nurse_Staffing #Institutional_Isomorphism #Bourdieu_Field_Theory #World_Systems_Theory #Health_Workforce References Cegłowska, U., Zawada, A., Zielińska, M., Hermanowski, T., & Buczacki, A. (2024). Using a lean management approach in acute ischemic stroke management: a systematic review. Postępy Psychiatrii i Neurologii, 33(2), 67–79. https://doi.org/10.5114/ppn.2024.141056 Chen, M., Guan, Q., & Zhuang, J. (2024). Patient-centered lean healthcare management from a humanistic perspective. BMC Health Services Research, 24(1), 1261. https://doi.org/10.1186/s12913-024-11755-w Collyer, F. M., Willis, K., Franklin, M., Harley, K., & Short, S. D. (2015). Healthcare choice: Bourdieu's capital, habitus and field. Current Sociology, 63(5), 685–699. https://doi.org/10.1177/0011392115590082 de Barros, L. B., de Camargo Bassi, L., Caldas, L. P., et al. (2021). Lean healthcare tools for processes evaluation: an integrative review. International Journal of Environmental Research and Public Health, 18(14), 7389. https://doi.org/10.3390/ijerph18147389 DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. Gherghina, Ș. C., et al. (2024). Lean management and predictive analytics in public hospital operations: empirical strategies for waste elimination and staffing optimisation. [Full bibliographic details to be confirmed by the author.] Giwa, and colleagues. (2025). Emerging dependencies: GCC states and Sub-Saharan African healthcare labor migration through world systems theory. World Medical & Health Policy. https://doi.org/10.1002/wmh3.70037 Hung, D. Y., Mujal, G., Jin, A., & Liang, S. Y. (2021). Patient experiences after implementing lean primary care redesigns. Health Services Research, 56(3), 363–370. https://doi.org/10.1111/1475-6773.13605 Ivcovici, A., Bamber, G., Bartram, T., Stanton, P., Borg, J., & Pariona-Cabrera, P. (2025). The regulation and reconciliation of hybrid professional-managerial identities in a public hospital: the case of Lean management. Journal of Health Organization and Management. https://doi.org/10.1108/JHOM-02-2024-0058 Lum, K. Y., Indera Putera, K. A. S., Krishnan, M., Libasin, Z., Abdullah, N. N. R., & Saman, I. S. (2023). The sustainability of lean implementations at the hospitals of Ministry of Health Malaysia: a study protocol. PLOS ONE, 18(11), e0294055. https://doi.org/10.1371/journal.pone.0294055 Marsilio, M., Pisarra, M., Rubio, K., & Shortell, S. (2022). Lean adoption, implementation, and outcomes in public hospitals: benchmarking the US and Italy health systems. BMC Health Services Research, 22, 122. https://doi.org/10.1186/s12913-022-07473-w McMahon, M., Plate, S., Herz, T., Brenner, G., Kleinknecht-Dolf, M., & Krauthammer, M. (2025). Development of a data-based method for predicting nursing workload in an acute care hospital: methodological study. Journal of Medical Internet Research, 27, e66667. https://doi.org/10.2196/66667 Meredith, P., Saville, C., Dall'Ora, C., Weeks, T., Wierzbicki, S., & Griffiths, P. (2025). Estimating nurse workload using a predictive model from routine hospital data: algorithm development and validation. JMIR Medical Informatics, 13, e71666. https://doi.org/10.2196/71666 Nover, J., Bai, M., Tismina, P., Raut, G., Patel, D., Nadkarni, G. N., Abella, B. S., Klang, E., & Freeman, R. (2025). Comparing machine learning and nurse predictions for hospital admissions in a multisite emergency care system. medRxiv (preprint). https://doi.org/10.1101/2025.04.07.25325126 Porto, B. M., & Fogliatto, F. S. (2024). Enhanced forecasting of emergency department patient arrivals using a feature engineering approach and machine learning. BMC Medical Informatics and Decision Making, 24, 377. Régis, T. K. O., Santos, L. C., & Gohr, C. F. (2019). A case-based methodology for lean implementation in hospital operations. Journal of Health Organization and Management, 33(6), 656–676. https://doi.org/10.1108/JHOM-09-2018-0267 Talero-Sarmiento, L. H., Escobar-Rodríguez, L. Y., Gomez-Avila, F. L., & Parra-Sanchez, D. T. (2024). A literature review on lean healthcare: implementation strategies, challenges, and future research directions. Cogent Engineering, 11(1), 2411857. https://doi.org/10.1080/23311916.2024.2411857 Trakulsunti, Y., & Trakoonsanti, L. (2021). The use of lean tools to reduce inpatient waiting time in a Thai public hospital: an action research study. Leadership in Health Services, 34(2), 84–97. https://doi.org/10.1108/LHS-10-2020-0080 Wallerstein, I. (2004). World-systems analysis: an introduction. Duke University Press. Wang, J., Lv, H., Chen, M., Liu, C., Ren, W., Jiang, H., & Zhang, L. (2025). A systematic review of lean implementation in hospitals: impact on efficiency, quality, cost, and satisfaction. International Journal of Health Policy and Management. https://doi.org/10.34172/ijhpm.8974

The emergence of #generative_AI, particularly #large_language_models (LLMs), is reshaping the landscape of #personalized_oncology and #clinical_decision_support. Grounded in the foundational diagnostic study by Benary et al. (2023), this article examines how LLMs are being integrated into the #precision_medicine workflow, enabling oncologists to navigate an increasingly complex terrain of #molecular_biomarkers, treatment guidelines, and patient-specific data. Drawing on Pierre Bourdieu's concepts of field and capital, world-systems theory, and the sociological framework of #institutional_isomorphism, this paper argues that the adoption of #AI_in_healthcare is not merely a technical transition but a deeply structured sociotechnical process shaped by institutional pressures, professional hierarchies, and global inequalities in access to advanced medical infrastructure. Through a narrative synthesis of recent empirical studies, this article analyses the performance, promise, and limitations of LLMs as decision-support tools in #tumor_board settings, #genomic_data interpretation, and personalized treatment matching. Findings suggest that while LLMs do not yet match the quality of human expert recommendations, they augment clinical reasoning, accelerate literature synthesis, and hold particular promise when deployed within retrieval-augmented generation (RAG) pipelines. Key barriers to adoption — including #AI_hallucination, data governance, equity of access, and the absence of standardized evaluation benchmarks — are discussed in depth. The article concludes with recommendations for responsible integration of #AI_tools in oncology practice. Keywords: generative artificial intelligence, large language models, oncology, clinical decision support, precision medicine, institutional isomorphism, Bourdieu, personalized treatment 1. Introduction Cancer remains one of the most diagnostically and therapeutically complex domains in modern medicine. Each patient presents a unique constellation of #tumor_genetics, co-morbidities, prior treatment histories, and psychosocial factors that resist simple, protocol-driven management. The rise of #precision_oncology over the past two decades has only deepened this complexity: oncologists must now interpret vast volumes of #next_generation_sequencing data, cross-reference rapidly evolving clinical guidelines, and identify off-label or trial-based options that may benefit individual patients — all within the compressed timescales of clinical consultations. Into this complexity, a new class of computational tool has arrived. #Large_language_models — artificial intelligence systems trained on enormous corpora of text and capable of generating coherent, contextually appropriate responses — have begun to attract serious attention from the oncology community. These systems, ranging from OpenAI's GPT-4 to open-source alternatives like Llama 2, can read, summarise, and reason across large bodies of biomedical literature, offering a potentially scalable path toward real-time #clinical_decision_support. The pivotal study underpinning this article is that of Benary et al. (2023), published in JAMA Network Open. That study evaluated four LLMs — ChatGPT, Galactica, Perplexity, and BioMedLM — using ten fictional patients with advanced cancers and genetic alterations. Each case was submitted to both the LLMs and a human expert, with #treatment_options then reviewed by a molecular #tumor_board. The study found that while LLMs deviated substantially from expert recommendations, they nonetheless identified several important treatment strategies and produced suggestions that were genuinely useful to the board. Crucially, at least one useful treatment option per patient was generated only by an LLM. This finding captures the essential tension at the heart of #AI_in_oncology: these tools are not ready to replace human clinical judgement, but they are already capable of augmenting it in meaningful ways. This article examines that tension through multiple lenses — technical, institutional, sociological, and ethical — to assess where #AI_decision_support is heading and what conditions are necessary for its responsible integration into clinical care. The article proceeds as follows. Section 2 establishes the theoretical framework, drawing on Bourdieu's field theory, world-systems analysis, and institutional isomorphism. Section 3 describes the methodological approach used in this narrative review. Section 4 analyses the current state of the evidence. Section 5 presents the key findings. Section 6 concludes with implications for practice and policy. 2. Background and Theoretical Framework 2.1 The Clinical Decision-Making Field in Oncology Pierre Bourdieu's concept of the field — a structured social space in which agents compete for forms of capital — offers a productive lens through which to understand #AI_integration in oncology. The oncological field is organised around multiple forms of capital: the scientific capital of those who produce cutting-edge research, the medical capital held by experienced clinicians at elite academic centres, and the institutional capital of hospitals and tumour boards that confer legitimacy on treatment decisions. The introduction of #AI_tools into this field does not occur in a neutral space; it disrupts existing hierarchies, challenges professional authority, and reconfigures what counts as valid knowledge. For Bourdieu, agents in a field engage in practices shaped by their habitus — the durable dispositions formed through prior socialisation. Senior oncologists trained in an era of purely expert-driven decision-making may resist the integration of AI not because they are unreasonable, but because such integration is at odds with their professional habitus. Conversely, junior clinicians and informaticists, whose habitus has been shaped by a digitally saturated training environment, may embrace these tools more readily. The uneven uptake of #AI_decision_support systems across different clinical settings and career stages reflects precisely this Bourdieuisian dynamic. 2.2 World-Systems Theory and Global AI Inequity Immanuel Wallerstein's world-systems theory draws attention to the structural inequalities between core, semi-periphery, and periphery nations in the global economy. Applied to #AI_in_healthcare, this framework reveals a troubling geography: the most sophisticated LLM-based #clinical_decision_support tools are being developed, validated, and deployed predominantly in high-income, technologically advanced countries — the United States, Germany, Japan, and the United Kingdom. The datasets on which these models are trained are largely drawn from patient populations in these same nations, meaning that their accuracy and clinical relevance may be systematically lower for patients in low- and middle-income countries where cancer burden is disproportionately high. From a world-systems perspective, the global diffusion of AI in oncology risks replicating — rather than reversing — structural inequities in #cancer_care. Peripheral nations that lack the computational infrastructure, curated clinical datasets, and regulatory capacity to develop their own systems may become dependent on imported tools that do not adequately reflect their patient populations, disease epidemiology, or healthcare contexts. This concern is not merely theoretical. As Benson et al. (2024) noted in their narrative review, applied LLM methods in cancer care still lag significantly behind advances in computer science, and standardised evaluation metrics are absent across the field. The gap between technological possibility and equitable deployment is a structural feature of how AI in medicine is currently organised. 2.3 Institutional Isomorphism and the Spread of AI Tools DiMaggio and Powell's concept of institutional isomorphism describes the process by which organisations come to resemble one another not because of efficiency gains but because of pressures to appear legitimate within a shared institutional field. Three mechanisms drive this convergence: coercive isomorphism (pressure from regulators and funders), mimetic isomorphism (copying what successful peers are doing), and normative isomorphism (adoption of practices endorsed by professional associations). All three mechanisms are visible in the spread of #AI_decision_support tools across oncology. Regulatory signals — such as the U.S. FDA's growing framework for AI-enabled medical devices and the European Union AI Act — exert coercive pressure. Institutional competition between academic cancer centres creates mimetic pressure: when Memorial Sloan Kettering Cancer Center or Charité Berlin integrates LLM-based tools, peer institutions face pressure to follow suit. Professional societies, through clinical guidelines and position statements, exert normative pressure. The result is a wave of adoption that may outrun clinical evidence, as institutions race to signal modernity rather than waiting for robust #randomized_clinical_trial validation. Currie's (2012) longitudinal study of electronic health records adoption in the UK National Health Service demonstrated precisely this pattern, showing that institutional isomorphic pressures led to adoption ahead of proven benefit, resulting in resistance and delayed implementation. There is a genuine risk that #LLM_integration in oncology follows a similar trajectory. 3. Method This article employs a narrative review methodology, synthesising findings from peer-reviewed empirical studies, systematic reviews, and high-quality preprints published between 2021 and 2026. The focus is on studies examining the application of LLMs and #generative_AI to clinical decision support in oncology, with particular attention to studies that evaluate model performance against expert or guideline standards. The search terms used to identify relevant literature included: "large language models oncology," "generative AI clinical decision support cancer," "LLM precision medicine," "GPT tumour board," and related variants. Priority was given to studies published in Scopus-indexed Q1 and Q2 journals, including JAMA Network Open, Nature Cancer, BMJ Oncology, JCO Clinical Cancer Informatics, Annals of Oncology, and Artificial Intelligence Review. Studies were selected if they included empirical evaluation of LLM performance, quantitative benchmarking against human expert decisions, or systematic analysis of LLM applications across cancer types. The theoretical framework sections draw on classical and contemporary works in sociology, institutional theory, and science and technology studies to contextualise findings within broader sociotechnical dynamics. Sociological theories were applied interpretively rather than formally coded, following established traditions of theoretically informed narrative review in medical sociology and health informatics. No primary data were collected for this study. Ethical approval was not required. 4. Analysis 4.1 LLMs as Decision-Support Tools: Performance Evidence The most methodologically rigorous evidence base for LLM performance in #clinical_decision_support comes from studies that directly compare model outputs with decisions made by expert oncologists or multidisciplinary tumour boards. The Benary et al. (2023) study established the benchmark. Across ten complex patient cases, the four LLMs achieved median F1 scores of 0.04, 0.17, 0.14, and 0.19 when compared against a human expert as the gold standard. When the outputs of all LLMs were combined, precision and recall each reached 0.29, suggesting that ensemble use of models offers modest improvement. Crucially, the molecular tumour board judged at least one LLM-generated treatment option per patient as useful, including novel strategies not identified by the human annotator. Subsequent research has built on and extended this baseline. Ferber et al. (2025), publishing in Nature Cancer, developed an autonomous AI agent integrating GPT-4 with precision oncology tools including vision transformers for microsatellite instability detection and web-based search capabilities. Evaluated on twenty realistic multimodal patient cases, the integrated system correctly used appropriate tools in 87.5% of cases, reached correct clinical conclusions in 91% of instances, and accurately cited oncology guidelines 75.5% of the time. Notably, GPT-4 alone achieved only 30.3% accuracy on the same cases, demonstrating that isolated LLM capability is insufficient and that performance rises dramatically when models are embedded within structured tool ecosystems. Studies examining LLM concordance with multidisciplinary tumour board decisions have largely found moderate-to-high agreement in well-defined clinical scenarios. Erdat et al. (2025), in a cross-sectional study published in Clinical and Translational Oncology, evaluated GPT-4 against 610 tumour board decisions at Ankara University Hospitals. The mean compatibility score between GPT-4 predictions and MTB decisions was 3.59 on a 4-point scale, with excellent inter-rater reliability. Discrepancies clustered around rare cancers and complex comorbid presentations, areas where current guidelines offer limited guidance and where the model's training data may be sparse. Büyükceran et al. (2025) reported similarly encouraging results in breast cancer specifically, finding that GPT-4o achieved full concordance with MTB decisions in 31 of 33 cases (93.9%) when provided with structured clinical input. The mean concordance score was 4.94 out of 5.0, with perfect inter-rater reliability. The two discordant cases involved questions about genetic counselling referral, suggesting that the model's limitations are not primarily in treatment selection but in the more nuanced, patient-centred dimensions of care planning. Less encouraging results emerged from the breast cancer benchmarking study by Shah et al. (2025/2026), a multi-institutional expert evaluation comparing three LLMs — GPT-4o, DeepSeek-R1, and OpenEvidence — across ten complex breast cancer scenarios rated by sixteen breast medical oncologists. OpenEvidence achieved the highest overall performance score of 3.91 out of 5, significantly outperforming both GPT-4o (3.19) and DeepSeek-R1 (2.93). GPT-4o showed the highest variability in expert ratings, reflecting inconsistent performance across cases. This variability is not a trivial concern: in clinical practice, an unpredictable tool is in some ways more dangerous than a consistently limited one, because clinicians cannot reliably calibrate when to override its suggestions. 4.2 Retrieval-Augmented Generation and Precision Oncology A consistent theme emerging from the evidence base is that #retrieval_augmented_generation (RAG) represents the most promising near-term pathway for LLM deployment in #precision_oncology. In a RAG pipeline, the LLM's responses are grounded in dynamically retrieved, trusted external knowledge sources — such as curated oncology databases, clinical guidelines, and peer-reviewed literature — rather than relying solely on patterns encoded in its weights during training. Jun et al. (2026), publishing in Cancer Cell, developed a RAG-LLM workflow using the Molecular Oncology Almanac (MOAlmanac) as a knowledge base and benchmarked it against an LLM-only approach. The RAG pipeline achieved up to 95% accuracy on synthetic queries and 93% accuracy on real-world queries from practising oncologists. By contrast, the LLM-only approach performed substantially worse. The authors concluded that this system could serve as actionable guidance for deploying LLMs in precision oncology clinical settings. Similar findings were reported by Kreimeyer et al. (2024), who demonstrated that a RAG pipeline using trusted treatment relationships from OncoKB reproduced over 80% of those relationships when queried using simple questions posed to an untrained Llama 2 model. This finding is significant because it demonstrates that even a general-purpose model can achieve high accuracy in oncological treatment matching if provided with reliable, structured context at inference time — without expensive fine-tuning on proprietary clinical datasets. For Hamamoto et al. (2026), in a comprehensive review published in the Journal of Hematology and Oncology, a RAG strategy combined with human-in-the-loop (HITL) workflows represents the most viable near-term model for responsible integration of #generative_AI in precision oncology. The authors describe a governance architecture centred on standardised clinical data models (OMOP, mCODE, FHIR), multicenter prospective evaluation, auditable decision logs, and synthetic data generation with differential privacy protections. This architecture is explicitly designed to address the #AI_governance challenges that have slowed clinical adoption. 4.3 Multimodal Applications and Tumour Board Support Cancer diagnosis and management is inherently multimodal, relying on integrated interpretation of pathology reports, radiology images, genomic sequencing outputs, clinical notes, and laboratory results. Early LLM applications in oncology were limited to text-based inputs, but the evidence base is rapidly expanding to include multimodal foundation models capable of processing images alongside text. Vasanthapuram (2025) described an AI pipeline combining generative models with LLMs for early cancer detection from radiological imaging, demonstrating improved diagnostic accuracy and reduced report generation time. Ferber et al. (2025) integrated vision transformers capable of detecting microsatellite instability, KRAS, and BRAF mutations from histopathology slides with language model reasoning about treatment implications, demonstrating that the combination substantially outperforms text-based reasoning alone. In tumour board settings specifically, Wang et al. (2025), publishing in iScience, synthesised evidence across cancer types and found that AI recommendations agreed with MDT decisions in 70–90% of cases, with the highest concordance observed for treatment intent, surgical decisions, and radiotherapy planning. Systemic therapy decisions showed the lowest AI-human agreement, reflecting the greater complexity and nuance of chemotherapy and immunotherapy selection. Nardone et al. (2024), reviewing AI's role in #multidisciplinary_tumor_boards from the perspectives of surgeons, medical oncologists, and radiation oncologists, found broad clinical enthusiasm for AI assistance in information synthesis, guideline retrieval, and documentation, but strong consensus that AI should remain advisory rather than autonomous. The fear of #clinician_deskilling — where repeated reliance on AI gradually erodes the independent diagnostic capabilities of oncologists — was raised as a significant concern by experienced clinicians. 4.4 Patient-Facing Applications and Health Literacy The impact of #AI_tools in oncology extends beyond the clinician's consultation room. Chen et al. (2025), in a narrative review published in BMJ Oncology, described the emerging category of patient-facing LLMs capable of disseminating accessible cancer information, answering questions about treatment options in plain language, and providing psychosocial support. This represents a qualitative democratisation of health information that aligns with broader commitments to #patient_empowerment and informed consent. However, the democratisation narrative must be held against its structural constraints. From a world-systems perspective, the ability to benefit from patient-facing AI is conditioned on digital literacy, device access, language coverage, and healthcare infrastructure. In peripheral nations and among marginalised populations in core nations, these conditions are unevenly distributed. #AI_generated health information that is accurate for patients in clinical trial settings may be misleading for those in settings where those treatments are unavailable or unaffordable. 5. Findings 5.1 LLMs Augment but Do Not Replace Clinical Expertise The strongest and most consistent finding across the studies reviewed is that LLMs function best as augmentative tools rather than autonomous decision-makers. No study found that any current LLM, deployed without structured human oversight, produced recommendations of sufficient quality and consistency to replace the expert judgement of a trained oncologist. Benary et al. (2023) established this finding clearly, and subsequent research has refined rather than reversed it. At the same time, the augmentative value is genuine. When LLMs are integrated into tumour board workflows, they accelerate the synthesis of relevant literature, surface treatment options that human annotators may overlook, and provide a first-pass organisation of complex multimodal data that reduces cognitive load for clinicians. Derbal (2025), framing cancer treatment as a problem of controlling a nonlinear dynamical system, argued that generative AI's capacity for data-driven prediction of disease state and treatment response dynamics positions it as a catalyst — not a replacement — for effective oncological decision-making. 5.2 RAG Architecture Is Superior to Standalone LLMs A clear and consistent finding is that standalone LLMs — queried without access to curated, current knowledge sources — perform substantially worse than RAG-augmented systems that ground responses in trusted oncology databases. This finding has direct implications for clinical deployment: institutions considering LLM integration should not deploy general-purpose chat systems directly into clinical workflows but should instead invest in RAG pipelines that connect language models to validated, regularly updated knowledge bases. The Woollie model, described by Heydari et al. (2025) in npj Digital Medicine, exemplifies this approach. Fine-tuned using radiology impression notes from a major cancer centre and externally validated on oncology datasets, Woollie demonstrates that domain-specific fine-tuning combined with curated training data can achieve both high accuracy and genuine scalability. The model prioritises data accuracy, guards against catastrophic forgetting, and establishes a foundation for equitable LLM deployment in oncology. 5.3 Hallucination and Bias Remain Unresolved Challenges #AI_hallucination — the generation of plausible-sounding but factually incorrect clinical information — is the single most serious technical limitation of current LLMs in clinical settings. In oncology, where an incorrect treatment recommendation can result in patient harm, hallucination is not merely a nuisance but a patient safety concern. Multiple studies reviewed here identified hallucination as a primary barrier to clinical adoption, and none reported a reliable solution. Bias is a related but distinct concern. LLMs trained predominantly on English-language, Western clinical literature will encode the demographic and epidemiological biases present in that literature. When deployed in settings with different patient populations, they may produce recommendations that are subtly but significantly less appropriate. Ghebrehiwet et al. (2024), reviewing generative AI in personalised medicine, identified the limitation of foundation models in digital diagnostics, particularly for patient populations underrepresented in training data. From a Bourdieuisian perspective, this is a form of symbolic violence: the imposition of medical knowledge produced in high-capital contexts as though it were universal, without acknowledgement of its situated origins. The patients most likely to be harmed by biased AI recommendations are those already most disadvantaged within the healthcare field. 5.4 Institutional Adoption Is Outpacing Evidence Driven by the three mechanisms of institutional isomorphism — coercive regulatory signals, mimetic competition between elite cancer centres, and normative endorsement from professional oncology societies — the institutional adoption of #AI_tools is proceeding at a pace that exceeds the available clinical evidence. As Wang et al. (2025) noted, most AI-MTB concordance studies are retrospective and conducted at single centres, limiting generalizability. Rigorous prospective, multicentre evaluation of LLMs in real oncology workflows remains rare. Kim et al. (2023), in a multi-organisational qualitative study of AI adoption in US healthcare systems, identified eight key decision points in the institutional AI adoption lifecycle, many of which relate to governance, transparency, and accountability. The study found that health system leaders frequently adopted AI technologies without adequate governance structures in place, replicating the pattern Currie (2012) identified in EHR implementation. For world-systems theory, this trajectory is particularly concerning for peripheral healthcare systems. When AI adoption in oncology is driven by institutional prestige rather than evidence, the technologies that spread globally are those developed and validated for core-nation patient populations, leaving periphery institutions to adopt tools whose safety and effectiveness for their patients has not been established. 5.5 Genomic Data Interpretation and Precision Matching One of the most promising specific applications of LLMs in oncology is the interpretation of #genomic_data for treatment matching. The complexity of next-generation sequencing reports — which may identify dozens of mutations, many of uncertain clinical significance — makes this an ideal domain for AI assistance: the volume of relevant literature exceeds what any clinician can stay current with, and the consequences of missing a clinically actionable variant are severe. Ding et al. (2024), presenting OncoGPT at the Journal of Clinical Oncology, demonstrated that an AI system integrating NGS data analysis with mutation-treatment matching algorithms achieved a weighted F1 score of 0.926 for tissue-of-origin classification across eight cancer types, and significantly improved survival outcomes (hazard ratio 0.326) for patients matched using the system. Chen et al. (2024), presenting the ACMA system, reported that LLM-based treatment matching using real-world data from over 100,000 clinical cases improved median survival rates and achieved a 90% relevance match for clinical trial enrolment. These results, while preliminary and drawn from conference abstracts rather than peer-reviewed studies, point toward a future in which #AI_systems become standard components of the precision oncology workflow. 6. Conclusion The integration of #generative_AI and #large_language_models into oncology is no longer a speculative future — it is a present-tense transformation whose contours are only beginning to come into focus. The evidence reviewed in this article supports a measured but genuinely optimistic assessment. LLMs cannot yet replace the clinical judgement of experienced oncologists, but they can meaningfully extend it: surfacing treatment options that might otherwise be missed, synthesising biomedical literature at scales beyond human capacity, and providing real-time decision support during tumour board deliberations. The Benary et al. (2023) study remains foundational in establishing both the promise and the ceiling of current LLM capability. Its finding that LLMs identified useful treatment options inaccessible to the human expert — while also producing recommendations of substantially lower overall quality — captures the essential paradox of #AI_augmentation in clinical medicine. It is precisely because LLMs can see different things, not better things, that their value as collaborative tools is real. The theoretical frameworks employed in this article illuminate dimensions of #AI_integration that purely technical analyses tend to obscure. Bourdieu's field theory reveals how the adoption of AI is structured by professional capital, habitus, and institutional hierarchy — and why resistance to these tools is not simply ignorance but a rational response to perceived threats to professional authority. World-systems theory forces attention to the global inequities that #AI_development in wealthy nations may reproduce or deepen. Institutional isomorphism explains why adoption is outpacing evidence, driven by legitimacy-seeking behaviour rather than demonstrated clinical benefit. For responsible integration, this article recommends five principles. First, LLMs in oncology should be deployed within RAG architectures that ground outputs in curated, validated knowledge bases, not as standalone chat systems. Second, all clinical LLM deployments should include human-in-the-loop oversight, with explicit governance frameworks that specify when and how AI recommendations can influence clinical decisions. Third, evaluation standards must be established and applied prospectively, across multiple centres and patient populations, before any LLM tool is considered ready for routine clinical use. Fourth, the training data and evaluation populations of oncological AI tools must be diversified to reduce the systematic biases that world-systems analysis identifies. Fifth, the healthcare community must resist the mimetic and coercive pressures of #institutional_isomorphism that encourage premature adoption, and insist that clinical safety evidence precede widespread deployment. The future of #AI_in_oncology is not the replacement of the oncologist, but the creation of a new kind of clinical partnership — one in which human expertise and machine capability are combined in ways that are greater than either alone. Getting that partnership right requires not only better models, but better institutions, better governance, and a more equitable global distribution of the technologies that will define cancer care in the coming decades. This is a call for science to be guided not only by what is technically possible, but by what is socially just. References Benary, M., Wang, X., Schmidt, M., Soll, D., Hilfenhaus, G., Nassir, M., Sigler, C., Knödler, M., Keller, U., Beule, D., Keilholz, U., Leser, U., & Rieke, D. (2023). Leveraging large language models for decision support in personalized oncology. JAMA Network Open, 6(11), e2343689. https://doi.org/10.1001/jamanetworkopen.2023.43689 Benson, R., Elia, M., Hyams, B., Chang, J., & Hong, J. C. (2024). A narrative review on the application of large language models to support cancer care and research. Yearbook of Medical Informatics, 33, 1–12. https://doi.org/10.1055/s-0044-1800726 Benson, R., Kenny, C., Ganjouei, A. A., Zhao, M., Darawsheh, R., Qian, A., & Hong, J. C. (2025). Large language models in population oncology: a contemporary review on the use of LLMs to support data collection, aggregation, and analysis in cancer care and research. JCO Clinical Cancer Informatics, 9, e2500112. https://doi.org/10.1200/CCI-25-00112 Büyükceran, E. U., Seyfettin, A., Babatürk, A., Özkan, M. B., Çolak, D., Ünal, İ., Kaymaz, E., Ergün, E., Emер, M. Ö., & Mersin, H. (2025). High concordance between GPT-4o and multidisciplinary tumor board decisions in breast cancer: a retrospective decision support analysis. Journal of Medical Systems, 49(1), 1–10. https://doi.org/10.1007/s10916-025-02314-9 Chen, D., Parsa, R., Swanson, K., Nunez, J. J., Critch, A., Bitterman, D., Liu, F. F., & Raman, S. (2025). Large language models in oncology: a review. BMJ Oncology, 4(1), e000759. https://doi.org/10.1136/bmjonc-2025-000759 Chen, H., Xu, Z., Zhang, P., Pang, F., Chen, L., Yuan, S., & Wang, K. (2024). Precision oncology and AI: the ACMA system for personalized treatment matching using real-world data. Journal of Clinical Oncology, 42(16 Suppl.), e23134. https://doi.org/10.1200/jco.2024.42.16_suppl.e23134 Currie, W. (2012). Institutional isomorphism and change: the national programme for IT – 10 years on. Journal of Information Technology, 27(3), 236–248. https://doi.org/10.1057/jit.2012.18 Das, U. (2025). Transforming precision medicine through generative AI: advanced architectures and tailored therapeutic design for patient-specific drug discovery. ChemistrySelect, 10(18), e202502448. https://doi.org/10.1002/slct.202502448 Derbal, Y. (2025). Generative artificial intelligence as a catalyst for effective cancer treatments. AI in Clinical Medicine, 3, 1–18. https://doi.org/10.14740/aicm5 Ding, S. D., Xing, X. M., Ding, Y., & Huang, Q. (2024). OncoGPT: an AI assistant for genomic-driven precision oncology. Journal of Clinical Oncology, 42(23 Suppl.), 160. https://doi.org/10.1200/jco.2024.42.23_suppl.160 Erdat, E. C., Yalçıner, M., Örüncü, M. B., Ürün, Y., & Senler, F. (2025). Assessing the accuracy of the GPT-4 model in multidisciplinary tumor board decision prediction. Clinical and Translational Oncology, 27(5), 1–10. https://doi.org/10.1007/s12094-025-03905-1 Ferber, D., El Nahhas, O. E., Wölflein, G., Wiest, I., Clusmann, J., Leßmann, M., Foersch, S., Lammert, J., Tschochohei, M., Jaeger, D., Salto-Tellez, M., Schultz, N., Truhn, D., & Kather, J. (2025). Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nature Cancer, 6(4), 621–634. https://doi.org/10.1038/s43018-025-00991-6 Ghebrehiwet, I., Zaki, N., Damseh, R., & Mohamad, M. S. (2024). Revolutionizing personalized medicine with generative AI: a systematic review. Artificial Intelligence Review, 57(6), 1–38. https://doi.org/10.1007/s10462-024-10768-5 Hamamoto, R., Koyama, T., Takahashi, S., Yasuda, T., Kobayashi, K., Akagi, Y., Kouno, N., Sudo, K., Hirata, M., Sunami, K., & Yamamoto, N. (2026). Implementing generative artificial intelligence in precision oncology: safety, governance, and significance. Journal of Hematology and Oncology, 19(1), 1–25. https://doi.org/10.1186/s13045-026-01781-y Heydari, K., Enichen, E. J., Li, P. C., & Kvedar, J. C. (2025). Incorporating large language models as clinical decision support in oncology: the Woollie model. npj Digital Medicine, 8(1), 1–5. https://doi.org/10.1038/s41746-025-01941-3 Jun, H., Tanaka, Y., Johri, S., Camp, S., Bao, E. L., Carvalho, F. L., Gui, D. Y., Jordan, A. C., Labaki, C., Martin, S. A., Nagy, M., O'Meara, T. A., Pappa, T., Pimenta, E., Saad, E., Yang, D. D., Gillani, R., Tewari, A., Reardon, B., & Van Allen, E. V. (2026). A context-augmented large language model for accurate precision oncology medicine recommendations. Cancer Cell, 44(2), 1–15. https://doi.org/10.1016/j.ccell.2025.12.017 Kehl, K. L. (2025). Use of large language models in clinical cancer research. JCO Clinical Cancer Informatics, 9, e2500027. https://doi.org/10.1200/CCI-25-00027 Kim, J., Boag, W., Gulamali, F., Hasan, A., Hogg, H., Lifson, M. A., Mulligan, D., Patel, M. R., Raji, I. D., Sehgal, A., Shaw, K., Tobey, D., Valladares, A., Vidal, D., Balu, S., & Sendak, M. (2023). Organizational governance of emerging technologies: AI adoption in healthcare. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1–18). ACM. https://doi.org/10.1145/3593013.3594089 Kreimeyer, K., Canzoniero, J., Fatteh, M., Anagnostou, V., & Botsis, T. (2024). Using retrieval-augmented generation to capture molecularly-driven treatment relationships for precision oncology. Studies in Health Technology and Informatics, 316, 1–7. https://doi.org/10.3233/SHTI240575 Nardone, V., Marmorino, F., Germani, M., Cichowska-Cwalińska, N., Menditti, V., Gallo, P., Studiale, V., Taravella, A., Landi, M., Reginelli, A., Cappabianca, S., Girnyi, S., Cwalinski, T., Boccardi, V., Goyal, A., Skokowski, J., Oviedo, R. J., Abou-Mrad, A., & Marano, L. (2024). The role of artificial intelligence on tumor boards: perspectives from surgeons, medical oncologists and radiation oncologists. Current Oncology, 31(9), 5021–5038. https://doi.org/10.3390/curroncol31090369 Pinninti, R., Gullapalli, R., Mallavarapu, K. M., Singareddy, R., Nekkanti, S. S., Nikhil, P., Shukla, S., Hariharan, N., Gudipudi, D., Suseela, K., Koppula, V., Rao, T., & Rajappa, S. (2026). Comparing artificial intelligence and multidisciplinary tumor board decision making in real-world cancer care: a prospective blinded concordance study. ESMO Real World Data and Digital Oncology, 3, e100688. https://doi.org/10.1016/j.esmorw.2026.100688 Rieke, D. T., Wang, X., Schmidt, M., Soll, D., Hilfenhaus, G., Nassir, M., Sigler, C., Knoedler, M., Keilholz, U., Leser, U., & Benary, M. (2023). AI-supported treatment decisions for precision oncology. Annals of Oncology, 34(Suppl. 4), S1536–S1536. https://doi.org/10.1016/j.annonc.2023.09.2869 Shah, Z., Afridi, S. S., Ombada, M., Roy, A. M., LeVee, A., Premji, S., Gupta, V., Lopetegui, N. M., Quiroga, D. M., Sacks, R. L., Shaikh, S., Abdou, Y., Yu, H., Madabhushi, A., Parikh, R., Chaudhary, L. N., Levine, E., Lambertini, M., Kalinsky, K., Kabraji, S., & Gandhi, S. (2025). Benchmarking large language models for clinical decision support in breast cancer care: a multi-institutional expert evaluation. Clinical Cancer Research, 32(4 Suppl.), Abstract PS3-04-06. https://doi.org/10.1158/1557-3265.sabcs25-ps3-04-06 Wang, X., Wang, Q., Ding, G., Wang, J., Tang, Y., & Feng, Y. (2025). Artificial intelligence in multidisciplinary tumor boards enhancing decision making and clinical outcomes in oncology. iScience, 28(6), 114082. https://doi.org/10.1016/j.isci.2025.114082 Yum, J., Naqvi, S. A. A., Zhou, B., & Riaz, I. (2025). Reimagining cancer care with generative artificial intelligence: the promise of large language models. JCO Clinical Cancer Informatics, 9, e2500134. https://doi.org/10.1200/CCI-25-00134 Hashtags #generative_AI #large_language_models #oncology #clinical_decision_support #precision_medicine #personalized_cancer_care #tumor_board_AI #AI_in_healthcare #molecular_oncology #LLM_oncology #retrieval_augmented_generation #AI_ethics_healthcare #institutional_isomorphism #AI_hallucination #cancer_genomics #AI_decision_making #multidisciplinary_oncology #digital_health_equity #Bourdieu_medicine #precision_treatment_matching

Health systems across the world are under strain from ageing populations, rising costs, climate pressure, and widening gaps between rich and poor patients. This article reviews how the field of #health_management is being redefined to meet those pressures. Building on the working definition offered by the European Health Management Association (EHMA, 2024), it argues that the old split between #public_health and #healthcare_administration no longer serves modern systems well. The review brings together recent peer-reviewed literature to describe a single, joined-up framework that treats prevention, service delivery, and stewardship of resources as parts of one task rather than separate jobs. To explain why systems behave the way they do, the analysis draws on three social theories: Pierre Bourdieu's ideas about #capital and #field, Immanuel Wallerstein's #world_systems theory, and the concept of #institutional_isomorphism developed by DiMaggio and Powell. Using a narrative-integrative method covering work published between 2019 and 2025, the article identifies four cross-cutting themes: governance and integration, #structural_efficiency, #sustainability and resilience, and #equity. The findings point to a framework with three layers (individual, organisational, and systemic) held together by shared data, mixed financing, and a workforce trained to think across boundaries. The article closes with implications for managers, educators, and policymakers, and notes where the evidence is still thin. Keywords: health management; public health; healthcare administration; equity; sustainability; institutional theory 1. Introduction For most of the twentieth century, the people who ran hospitals and the people who protected community #health worked in separate worlds. Hospital and clinic leaders dealt with budgets, staffing, beds, and billing. Public health officials dealt with outbreaks, sanitation, vaccination, and prevention. The two groups trained in different schools, read different journals, and reported to different parts of government. That division made some sense when most serious illness was acute and short-lived. It makes far less sense now. Today the heaviest burden on health systems comes from long-term conditions such as diabetes, heart disease, depression, and dementia. These conditions are shaped less by what happens inside a clinic and more by housing, income, food, work, and education, the so-called #social_determinants of health (Izadisabet, Aminbeidokhti, & Jafari, 2024). A patient with poorly controlled diabetes may visit a hospital dozens of times, yet the real drivers of their illness sit far upstream. Managing such a patient well requires a person who can read a balance sheet and also understand a neighbourhood. That is the person the modern era needs, and that is the role this article calls a #health_manager. The European Health Management Association has tried to capture this shift in a single definition. EHMA (2024) describes health management as the work of providing leadership and guidance to look after #health at the level of the individual, the organisation, and the whole system. In their view, health management holds a broad picture of health, one shaped by behaviour, social conditions, and the environment, and it stretches beyond the narrower task of #healthcare_management, which covers community, primary, secondary, and tertiary care. In plain terms, EHMA is saying that running a hospital is part of health management, but health management is bigger than running a hospital. This article reviews the literature that supports and challenges that claim. The central question is simple to state and hard to answer: can the principles of #public_health and the tools of #healthcare_administration be brought together into one coherent framework, and if so, what would that framework need to deliver #structural_efficiency, #sustainability, and #equity at the same time? Answering this question matters for three reasons. First, money is tight. Most wealthy countries already spend close to or above ten percent of national income on health, and the share is projected to keep climbing (OECD, 2024). Systems cannot keep paying for waste. Second, the climate and resource crisis is forcing every large sector, including health, to think about its own footprint and its own long-term survival (World Economic Forum, 2023). Third, the COVID-19 pandemic exposed in brutal detail how unequal access to care can be, both within and between countries (Cranfield, Li, & John-Schuster, 2024). A framework that improves efficiency but worsens fairness would fail the modern test. So would one that is fair but bankrupts the system. The article proceeds as follows. Section 2 sets out the theoretical framework, combining the EHMA definition with three sociological lenses. Section 3 explains the review method. Section 4 analyses the literature around four themes. Section 5 presents the unified framework that emerges from that analysis. Section 6 concludes with implications and limits. 2. Background and Theoretical Framework 2.1 From two fields to one field The first task is conceptual. Why have public health and healthcare administration stayed apart for so long, and what would joining them actually mean? Part of the answer is historical. The two grew from different roots, one in sanitary reform and epidemiology, the other in business administration and hospital operations. Part of the answer is structural. Each field built its own training programmes, professional bodies, funding streams, and status markers. To understand how these structures hold, and how they might be reshaped, this review uses three theories that each explain a different piece of the puzzle. 2.2 Bourdieu: capital, field, and the quiet reproduction of inequality Pierre Bourdieu (1986) gives us a vocabulary for thinking about how advantage is passed on. He argued that people carry different kinds of #capital. Economic capital is money and property. #social_capital is the value of one's networks and connections. #cultural_capital is the knowledge, habits, and credentials that signal belonging, including comfort with how institutions work. Bourdieu also used the idea of a #field, a structured space in which players compete using the capital they hold, and the idea of #habitus, the deep set of dispositions that shape how people act without thinking. Recent health research has put these ideas to work. Studies of older adults show that the mix of capitals a person holds, not just their income, predicts how and when they use services. People rich in social and cultural capital but poor in money tend to lean on primary care, while those with money lean on hospitals (Paccoud, Nazroo, & Leist, 2020). Comparative work across China and Japan finds that capital, habitus, and health form a tight triangle, so that #health_inequalities are reproduced quietly across generations through everyday choices and constraints (Lin, 2022). For #health_management, Bourdieu carries two lessons. First, equity cannot be solved by opening the clinic door alone, because patients arrive with very different capital and the same service does not land the same way for everyone. Second, the management profession itself is a field with its own forms of capital, which helps explain why public health knowledge and administrative knowledge have been treated as separate currencies rather than a shared one. 2.3 World-systems theory: core, periphery, and global health Bourdieu works well within a country. To explain patterns between countries, this review turns to Immanuel Wallerstein's (2004) #world_systems theory. Wallerstein described the world economy as a single system split into a wealthy #core, a poor #periphery, and an in-between #semi_periphery. Wealth and skilled labour flow toward the core, while the periphery supplies cheaper labour and raw goods, and the gap tends to persist over time. Applied to health, this lens helps explain stubborn global imbalances. Medical staff trained in lower-income countries migrate to richer ones. Research agendas, drug pricing, and the rules of #global_health_governance are largely set in the core. During COVID-19, vaccines reached core countries first while peripheral ones waited (Cranfield et al., 2024). A modern definition of health management that takes equity seriously cannot stop at national borders. It has to notice that a hospital in one country may be staffed, in part, by drawing talent away from another, and that sustainability for the rich can rest on extraction from the poor. 2.4 Institutional isomorphism: why organisations come to look alike The third lens explains a different puzzle: why do health organisations across very different settings end up looking so similar? DiMaggio and Powell (1983) called this #institutional_isomorphism. They identified three pressures. #coercive_pressure comes from laws, regulators, and funders who require certain structures. #mimetic_pressure leads organisations facing uncertainty to copy others they see as successful. #normative_pressure spreads through professions, as managers trained the same way bring the same playbook everywhere. These ideas remain central to current health-organisation research. Studies of digital transformation show hospitals adopting similar electronic systems partly because rivals do and regulators expect it, even when the fit is poor (Kjekshus et al., 2024). Work on #integrated_care finds organisations juggling several #institutional_logics at once, the logic of the market, of the profession, and of the state, which pull in different directions and slow real change (Roth & Vakkuri, 2024). The lesson for a unified framework is sobering. Simply telling organisations to integrate public health and #administration will not work if the surrounding pressures still reward the old separation. The framework has to change the incentives, not only the org chart. 2.5 Bringing the three lenses together Read together, the three theories describe the same problem at three scales. Bourdieu shows how #inequality reproduces inside a society through unequal capital. World-systems theory shows how it reproduces between societies through core and periphery. Institutional isomorphism shows how organisations resist deep change while appearing to modernise. A framework for modern health management has to push against all three currents at once. This is the theoretical backbone of the review that follows. 3. Method This article is a narrative-integrative literature review. That method suits a topic that crosses several fields and aims to build a framework rather than measure a single effect. The review followed four steps. Search. The author searched Scopus, PubMed, and Web of Science for peer-reviewed work in English. Search terms combined health management, healthcare administration, and public health with concepts such as integration, governance, #efficiency, sustainability, resilience, and equity, and with the names of the three guiding theories. The grey literature from major bodies such as the OECD and EHMA was included where it shaped the academic debate. Time window. The review prioritised work published between 2019 and 2025 so that the picture reflects the post-pandemic landscape. Foundational theory texts that fall outside this window, namely the core statements of Bourdieu, Wallerstein, and DiMaggio and Powell, were kept because the newer studies build directly on them and cannot be read without them. Selection. Studies were included if they spoke to the relationship between public health and #management, or applied one of the three theories to a health setting, or addressed at least one of the three target outcomes of #efficiency, sustainability, and equity. Purely clinical studies with no management or system angle were set aside. Opinion pieces without evidence were used only for context. Synthesis. Rather than pooling numbers, the review used thematic synthesis. The author read each source, drew out its main claims, and grouped these into recurring themes. The themes were then read back against the three theories, a step sometimes called theoretical triangulation, to check whether the patterns in the evidence matched what the theories would predict. Four themes held up across the literature and structure the next section: governance and #integration, structural efficiency, sustainability and resilience, and equity. The method has limits, which Section 6 returns to. It is a single-author review, so selection involves judgement. It leans on English-language sources, which under-represents the very #periphery that world-systems theory tells us to watch. These limits do not undo the findings, but they shape how far the findings can be pushed. 4. Analysis 4.1 Theme one: governance and integration The clearest message in the recent literature is that #governance, not technology or money, is the binding constraint. EHMA (2024) treats #governance as the frame through which systems set goals, oversee quality, and stay accountable, and it notes that no single agreed model of health-system governance yet exists. That gap is the opening that a unified framework tries to fill. Several studies show what good #integration looks like and why it is hard. Research on integrated care describes three layers that have to line up: structural integration (merging organisations and budgets), institutional integration (aligning rules and logics), and relational integration (building trust between people who used to work apart). Many reforms manage the first and ignore the other two, which is why merged organisations often keep working in their old #silos (Roth & Vakkuri, 2024). The institutional-logics literature explains the mechanism. When the market logic, the professional logic, and the public-service logic all operate at once, staff spend energy managing the clash rather than serving patients (Kjekshus et al., 2024). Through a Bourdieusian lens, each logic is a different currency in the same field, and integration fails when no one agrees on the exchange rate. 4.2 Theme two: structural efficiency Efficiency in health is widely misunderstood as simply cutting cost. The stronger literature treats structural efficiency as getting more health per unit of resource across the whole system, not just inside one unit. A rapid evidence synthesis of efficiency improvement at the system level found that the biggest gains come from better coordination, smarter use of data, and reducing duplication, rather than from squeezing front-line staff (Walters, Sharma, Malica, & Harrison, 2022). In other words, the waste lives in the gaps between organisations, exactly the gaps a unified framework is meant to close. This is also where institutional isomorphism bites. Hospitals often adopt expensive systems because peers and regulators expect them to, a #mimetic_pressure, even when a cheaper route would serve patients better (Kjekshus et al., 2024). True efficiency therefore depends on managers who can resist copying for its own sake and judge what their own population actually needs. That capacity is partly a question of #cultural_capital within the profession: knowing the field well enough to break from it wisely. 4.3 Theme three: sustainability and resilience After COVID-19, sustainability in health has come to mean two things at once. The first is financial and ecological staying power, keeping the system affordable and lowering its environmental footprint. The second is #resilience, the ability to absorb shocks and keep working. The OECD's renewed performance framework now treats efficiency, equity, sustainability, and resilience as cross-cutting dimensions that must be judged together rather than traded off one by one (OECD, 2024). The Partnership for Health System Sustainability and Resilience makes a similar case, arguing that a system is only sustainable if it can also bend without breaking under pressure (World Economic Forum, 2023). A systems-thinking review of public-health programmes adds a practical point. Programmes survive over the long run when they are designed with feedback, local ownership, and adaptive capacity from the start, not bolted on later (Nafees, AboAlfa, & Alkhaldi, 2025). This matches the institutional view: durable change requires shifting the underlying #institutional_logics, not just launching a project. It also matches the world-systems view, since periphery systems that depend on outside funding lose resilience the moment that funding moves. 4.4 Theme four: equity Equity is the theme where the three theories speak most loudly together. The evidence is now strong that #social_determinants drive most of the variation in #health_outcomes, and that addressing them requires data and partnerships reaching well beyond the clinic (Li, Guo, Bian, & Becich, 2025; Craven et al., 2024). Yet integrating that data is technically and politically hard, and uneven data collection itself reflects and reinforces existing #inequality (Craven et al., 2024). Bourdieu explains why a one-size service deepens gaps: patients with less capital extract less benefit from the same offer (Paccoud et al., 2020; Lin, 2022). World-systems theory explains why equity has a global face: workforce drains and unequal access track the #core_periphery line (Cranfield et al., 2024). And isomorphism warns that equity goals can become hollow ritual, adopted to satisfy funders while practice stays unchanged, a #normative_pressure that produces the appearance of fairness without the substance. A credible framework has to measure equity in outcomes, not in mission statements. 4.5 Reading the themes against the theory Across all four themes the same pattern appears. The technical fixes are reasonably well understood. The barriers are social and structural: unequal capital, global hierarchy, and organisations that change their surface faster than their core. This is the gap the proposed framework is built to address. 5. Findings: A Unified Framework for Modern Health Management Pulling the analysis together, the review yields a framework with three levels and three connective tissues. It restates and extends the EHMA (2024) definition by spelling out what the joined-up practice requires in concrete terms. 5.1 Three levels of action The individual level. Here health management means helping each person navigate care and protect their own health, with explicit attention to the different capital they bring. Practically, this means meeting patients where they are: plain-language communication, help with the social side of illness, and support for those with low cultural capital so that the same service produces fairer results (Paccoud et al., 2020). The organisational level. Here the task is to run services that are efficient and humane at once. The finding is that structural efficiency comes from coordination across organisations, shared data, and the courage to skip costly copying that does not fit local need (Walters et al., 2022; Kjekshus et al., 2024). Managers at this level need both administrative and public health literacy, the very blend the old training split denied them. The systemic level. Here health management becomes stewardship of the whole, setting policy, financing, and the rules of governance so that prevention and care reinforce rather than compete. This is where sustainability and #resilience are won or lost (OECD, 2024; World Economic Forum, 2023), and where global equity must be considered, since one system's gain can be another's loss (Cranfield et al., 2024). 5.2 Three connective tissues The three levels only become one system when three things bind them. First, shared data. Information on social determinants and on outcomes must flow across organisations and sectors, with common standards, so that the system can see itself whole (Li et al., 2025; Craven et al., 2024). Second, blended financing. Money has to reward keeping people well, not only treating them when sick. Payment models that pool funds across prevention and care reduce the #mimetic and #coercive pressures that lock organisations into the old split (Roth & Vakkuri, 2024). Third, a boundary-spanning workforce. The framework lives or dies on people trained to work across the line between public health and #administration. This is the deepest change, because it rewrites the cultural capital of the profession itself and, over time, softens the institutional isomorphism that keeps the two fields apart. 5.3 What the framework is built to resist The framework is not neutral. It is designed against the three forces the theory identified. Against Bourdieu's quiet reproduction of inequality, it pushes resources and design toward those with the least capital. Against the world-systems #core_periphery split, it asks systems to count the global cost of their own staffing and supply choices. Against institutional isomorphism, it tries to change incentives and training so that integration becomes the path of least resistance rather than a slogan layered over unchanged practice. 5.4 A note on what the evidence does not yet show Honesty requires flagging the gaps. The literature is strong on why integration matters and weaker on long-run proof that integrated systems outperform separated ones on cost and outcomes together. Much of the equity evidence describes the problem better than it tests solutions. And the global, world systems dimension is the least studied of all, partly because the research itself is produced mostly in the #core. The framework should therefore be read as a well-grounded proposal, not a settled result. 6. Conclusion The modern era has dissolved the old wall between public health and healthcare administration. Chronic disease, tight budgets, climate pressure, and the hard lessons of the pandemic all point the same way: prevention, care, and the stewardship of scarce resources are one task, not three. EHMA (2024) captured this when it defined health management as leadership for health at the individual, organisational, and systemic levels, broader than the management of care alone. This review has tried to give that definition analytical teeth. The three theories used here explain why the joining is so hard. Bourdieu shows how unequal capital reproduces #health_inequalities inside a society. Wallerstein's world systems theory shows how the core periphery divide reproduces them between societies. DiMaggio and Powell's institutional isomorphism shows how organisations can look modern while their core stays the same. A framework that wants structural efficiency, sustainability, and equity together must work against all three currents. The framework proposed in the findings does that through three levels of action, individual, organisational, and systemic, bound by shared data, blended financing, and a boundary-spanning workforce. For managers, the message is that efficiency and equity are not rivals but partners, since most waste hides in the gaps that also hurt the poorest patients. For educators, the message is that the next generation of leaders needs training that crosses the old line from day one. For policymakers, the message is that governance and incentives, not exhortation, are what finally make integration stick. The agenda ahead is clear. The field needs long-term studies that test whether integrated health management delivers on its promise, equity research that tests solutions and not only describes problems, and far more work produced in and about the periphery. If those gaps are filled, the unified framework can move from a reasoned argument to a proven practice. Until then it stands as the most coherent answer we have to a simple question: what does it mean to manage health well in the modern era? Hashtags #health_management #public_health #healthcare_administration #structural_efficiency #sustainability #equity #institutional_isomorphism #world_systems_theory #social_determinants #integrated_care #health_governance #resilience #cultural_capital #health_systems #population_health #HealthManagement #Health_Management_Modern_Era · #PublicHealthXAdministration · #EquityEfficiencySustainability · #EHMA2024 · #HealthSystemsReform · #DefiningHealthManagement · #BourdieuInHealth · #CorePeripheryHealth · #IsomorphismInHealthcare · #STULIB · #ScopusStyleReview · #HealthcareLeadership References Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of Theory and Research for the Sociology of Education (pp. 241–258). Greenwood Press. Cranfield, B., Li, G., & John-Schuster, G. (2024). 2024 at BMC Global and Public Health: a year in review. BMC Global and Public Health, 2, 78. https://doi.org/10.1186/s44263-024-00119-5 Craven, C. K., Highfield, L., Basit, M., Bernstam, E. V., Choi, B. Y., Ferrer, R. L., Gelfond, J. A., Pruitt, S. L., Kannan, V., Shireman, P. K., Spratt, H., Torres Morales, K. J., Wang, C.-P., Wang, Z., Zozus, M. N., Sankary, E. C., & Schmidt, S. (2024). Toward standardization, harmonization, and integration of social determinants of health data: A Texas Clinical and Translational Science Award institutions collaboration. Journal of Clinical and Translational Science, 8, e17. https://doi.org/10.1017/cts.2024.2 DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101 European Health Management Association. (2024). EHMA 2024 Conference Report: Shaping and managing innovative health ecosystems. European Health Management Association. Izadisabet, F., Aminbeidokhti, A., & Jafari, S. (2024). Social determinants of health during and after coronavirus: a qualitative study. BMC Public Health, 24, 283. https://doi.org/10.1186/s12889-024-17785-7 Kjekshus, L. E., et al. (2024). Competing institutional logics in healthcare organizations: theorising digitalism. Journal of Health Organization and Management, 38(8), 1299–1315. https://doi.org/10.1108/JHOM-03-2024-0107 Li, C., Guo, J., Bian, J., & Becich, M. J. (2025). Advancing social determinants of health research and practice: Data, tools, and implementation. Journal of Clinical and Translational Science, 9, e58. https://doi.org/10.1017/cts.2024.657 Lin, S. (2022). From developing to developed: Mechanisms of health inequalities among seniors in China and Japan under macro-field control. Frontiers in Psychology, 13, 956165. https://doi.org/10.3389/fpsyg.2022.956165 Nafees, Z., AboAlfa, M., & Alkhaldi, M. (2025). Operationalizing systems thinking approach to sustain public health rehabilitation programs: a rapid review and strategic synthesis. Frontiers in Rehabilitation Sciences, 6, 1633596. https://doi.org/10.3389/fresc.2025.1633596 Organisation for Economic Co-operation and Development. (2024). Rethinking Health System Performance Assessment: A Renewed OECD Framework. OECD Publishing. Paccoud, I., Nazroo, J., & Leist, A. (2020). A Bourdieusian approach to class-related inequalities: the role of capitals and capital structure in the utilisation of healthcare services in later life. Sociology of Health & Illness, 42(3), 510–525. https://doi.org/10.1111/1467-9566.13028 Roth, M., & Vakkuri, J. (2024). Value creation mechanisms in a social and health care innovation ecosystem – an institutional perspective. Journal of Management and Governance, 28(4), 1017–1048. https://doi.org/10.1007/s10997-024-09696-x Wallerstein, I. (2004). World-Systems Analysis: An Introduction. Duke University Press. Walters, J. K., Sharma, A., Malica, E., & Harrison, R. (2022). Supporting efficiency improvement in public health systems: a rapid evidence synthesis. BMC Health Services Research, 22, 293. https://doi.org/10.1186/s12913-022-07694-z World Economic Forum. (2023). Building Sustainable and Resilient Health Systems (Partnership for Health System Sustainability and Resilience). World Economic Forum.

As #genomic_medicine moves from research laboratories into routine #public_healthcare, the ability of ordinary people to understand and act on genetic information has become a public health priority rather than a niche concern. This article asks a forward-looking question: what strategic practices will populations and health systems need over the next decade to raise #genetic_literacy at scale? Drawing mainly on Kaphingst (2023) and a focused body of literature published within the past five years, the paper synthesises evidence using three social theories. Bourdieu's idea of #cultural_capital explains why genetic knowledge is unevenly distributed across social groups. #World_systems_theory explains why genomic capacity is concentrated in wealthy "core" nations while poorer regions remain dependent. #Institutional_isomorphism explains why health systems often copy each other's literacy programmes without adapting them to local need. The method is a structured narrative synthesis of recent peer-reviewed work. The analysis shows that low literacy is not simply a personal deficit but a structured outcome of unequal resources, global hierarchy, and imitative policy-making. The findings propose a set of strategic practices: standard measurement, school-to-adult learning pathways, workforce expansion, culturally grounded #health_communication, regulation of #direct_to_consumer_testing, and South-led capacity building. The article concludes that #future_forecasting in this field must treat literacy as an equity issue built into the design of #genomic_medicine, not an afterthought. 1. Introduction For most of the twentieth century, genetic information reached the public through a narrow channel. A patient with a family history of disease might be referred to a specialist, sit with a #genetic_counselor, and receive carefully framed advice. That model is fading. Today, genetic and genomic information reaches far larger segments of the population through clinical care, population screening programmes, and commercial testing kits sold straight to consumers (Kaphingst, 2023). The intermediary who once translated complex results is no longer guaranteed to be present. This shift creates an obvious problem. When people can access #genomic_data without a trained guide, their own knowledge and skills become the deciding factor in whether that data helps or harms them. Kaphingst (2023) calls this set of abilities #genetic_literacy, and argues that it is now central to whether genomic technologies actually improve health. The evidence is not reassuring. Studies repeatedly find large gaps in genetic understanding across the general population, and these gaps limit the benefit that #genomic_medicine can deliver (Kaphingst, 2023; Daly & Kaphingst, 2023). The stakes rise further as #precision_medicine becomes ordinary. #Polygenic_risk_scores, expanded carrier screening, pharmacogenomic testing, and #direct_to_consumer_testing are no longer rare events. A person may receive a risk estimate that is probabilistic, easy to misread, and emotionally heavy. Without adequate #numeracy and #health_literacy, that person may over-react, under-react, or make a costly decision based on a misunderstanding (Miyoshi, Watanabe, & Ayub, 2023). The communication challenge is real because the concepts are abstract and the information is rarely a simple yes or no (Daly & Kaphingst, 2023). The purpose of this article is not to describe the present alone. It is to forecast. #Future_forecasting, in the sense Kaphingst (2023) uses it, means identifying the practices that research and health systems should put in place now so that #genetic_literacy keeps pace with the spread of genomic tools. The article treats this as a strategic question that crosses three levels at once: the individual, the national health system, and the global order. To make sense of these levels, the paper uses three social theories rather than one. First, Bourdieu's theory of #cultural_capital helps explain why genetic knowledge clusters in some social groups and not others. Second, #world_systems_theory helps explain why genomic capacity is concentrated in a small number of wealthy nations while much of the world remains on the periphery. Third, #institutional_isomorphism helps explain why health systems tend to imitate one another's literacy programmes, sometimes adopting models that do not fit local realities. Together these lenses move the discussion beyond the common assumption that low literacy is just a matter of individual ignorance that more pamphlets will fix. The article is organised as follows. The next section sets out the theoretical framework and defines #genetic_literacy more precisely. The method section explains how the literature was gathered and synthesised. The analysis applies the three theories to the current state of #genomic_medicine. The findings section translates that analysis into concrete strategic practices. The conclusion considers what success would look like and where the risks lie. The central claim is straightforward: raising #genetic_literacy is achievable, but only if it is designed as an #health_equity project from the start, not bolted on once the technology has already widened existing gaps. 2. Background and Theoretical Framework 2.1 What genetic literacy means The term genetic literacy sounds simple but is contested. A systematic review by Daly and Kaphingst (2023) found that researchers define and measure it in many different ways, which makes results hard to compare across studies. Some measures test factual knowledge of genes and inheritance. Others test the ability to read and reason about risk numbers, which overlaps with #numeracy. Still others test the broader skill of finding, judging, and using genetic information in real decisions, which overlaps with #health_literacy more generally. For the purposes of this article, genetic literacy is treated as a layered ability. At its base sits factual knowledge: what a gene is, how traits are inherited, what a mutation does. Above that sits numeracy: the capacity to interpret probabilities, relative risks, and population frequencies. At the top sits what Abel and Benkert (2022) call critical literacy: the ability to question, contextualise, and act on information rather than simply receive it. A person who can recite genetic facts but cannot judge whether a #direct_to_consumer_testing result is meaningful has knowledge without critical literacy. Strategic practice must target all three layers. 2.2 Bourdieu: genetic knowledge as cultural capital Pierre Bourdieu argued that societies distribute not only money but also #cultural_capital, the knowledge, skills, tastes, and credentials that help people move through institutions with confidence (Bourdieu, 1986). Cultural capital is unevenly held. Families with higher education pass on habits of reading, questioning, and engaging with experts that Bourdieu called the #habitus. These habits give some people an advantage when they enter a clinic or read a complex test report. Recent health research confirms that this matters. Luo, Zhao, Chen, and Xiao (2023) found that #cultural_capital was positively associated with health literacy among community residents, partly through #social_capital. Paccoud, Nazroo, and Leist (2022) showed that the ability to navigate and optimise healthcare differs by social position and birthplace, reflecting unequal capital rather than unequal need. Applied to genomics, the implication is direct. #Genomic_data does not arrive in a social vacuum. People with more cultural capital are better placed to understand a #polygenic_risk_score, ask the right follow-up questions, and demand appropriate care. Without deliberate effort, genomic medicine will reward those who already hold the most capital and widen the very #health_disparities it promises to close (Halbert, 2022). This is the first reason that literacy is an equity problem. If genetic knowledge behaves like cultural capital, then simply releasing more information into the public sphere will benefit the already-advantaged fastest. Strategic practice must therefore redistribute genetic understanding, not just broadcast it. 2.3 World-systems theory: the global geography of genomic capacity If Bourdieu explains inequality within a society, #world_systems_theory explains inequality between societies. Immanuel Wallerstein described the modern world as a single economic system divided into a wealthy, technology-rich "core," a dependent "periphery," and an in-between "semi-periphery" (Wallerstein, 2004). Core nations concentrate high-value production and set the rules; peripheral nations supply raw inputs and remain dependent on the core for advanced capacity. Genomics maps onto this structure with unsettling neatness. A small number of core nations hold most of the sequencing infrastructure, reference databases, trained workforce, and intellectual property. The World Health Organization's Science Council warned in 2022 that low- and middle-income countries face shortfalls in financing, laboratory infrastructure, materials, and trained personnel, and argued that delayed access for poorer countries is neither ethically nor scientifically defensible (World Health Organization Science Council, 2022). Much of the world's genetic data, meanwhile, is drawn from populations of European ancestry, which means #polygenic_risk_scores and other tools often perform worse for everyone else. In world systems theory terms, peripheral regions may contribute samples, participants, and even raw genomic material while the analytic value, the products, and the literacy resources accumulate in the core. This is a genomic version of what Wallerstein called unequal exchange. The literacy consequence is severe: a population cannot easily build #genomic_literacy when the institutions, curricula, and professionals that teach it are concentrated elsewhere. #Future_forecasting that ignores this global hierarchy will simply forecast a wider gap. 2.4 Institutional isomorphism: why health systems copy each other The third lens explains a quieter problem. DiMaggio and Powell (1983) observed that organisations within a shared field tend to grow more alike over time, a process they called #institutional_isomorphism. They identified three pressures. Coercive pressure comes from regulation and funding conditions. Mimetic pressure comes from copying admired peers when goals are uncertain. Normative pressure comes from shared professional training that spreads common templates. In their later reflection, Powell and DiMaggio (2023) noted that these pressures remain powerful in fields marked by uncertainty, which describes genomic medicine well. For genetic literacy, isomorphism cuts both ways. On the positive side, it can spread good practice quickly: when one national health service adopts a respected literacy programme, others may follow. On the negative side, it encourages health systems to import literacy models, measurement tools, and #genetic_counseling structures designed for very different populations. A leaflet that works in one context is copied into another where reading levels, languages, and trust in institutions differ sharply. The result can be #policy_diffusion that looks modern and legitimate but does not actually raise understanding on the ground. Strategic practice must harness the spread of good ideas while resisting the unthinking imitation that isomorphism encourages. 2.5 Bringing the three lenses together These theories are not rivals; they describe three layers of the same problem. Bourdieu explains why literacy is unequal between people. World systems theory explains why it is unequal between nations. Institutional isomorphism explains why the remedies often fail to fit. A complete strategy for genetic literacy must answer all three at once: redistribute knowledge within societies, rebalance capacity between societies, and adapt rather than copy the tools used to teach. 3. Method This article is a structured narrative synthesis rather than an empirical study or a formal systematic review. The aim was to assemble a current, theory-informed account of how genetic literacy can be strengthened as genomic medicine becomes routine, and to do so using sources published recently enough to reflect the present state of the field. The starting point was the position paper by Kaphingst (2023), which frames future forecasting for research and practice in genetic literacy and was treated as the anchor text. From there, related literature was identified by following the themes Kaphingst raises: definition and measurement of literacy, #health_communication, #public_health_genomics, direct to consumer testing, and #health_equity. Priority was given to peer-reviewed articles and authoritative reports published within roughly the last five years, so that the synthesis would describe current rather than historical practice. Foundational theoretical texts by Bourdieu, Wallerstein, and DiMaggio and Powell were retained because they define the frameworks used, and a recent restatement of institutional theory (Powell & DiMaggio, 2023) was added to keep the theoretical discussion current. Sources were selected for relevance to three questions. First, what does the evidence say about the level and distribution of genetic literacy in populations? Second, what social and structural forces shape that distribution? Third, what interventions or policies have been tried or proposed? Systematic reviews were favoured where available, because they summarise many primary studies; examples include reviews of direct to consumer testing consequences (Nolan & Ormondroyd, 2023; Van Steijvoort et al., 2025) and of digital education interventions (Gasteiger et al., 2023). The synthesis proceeded in two steps. In the analysis step, the three theories were applied to the assembled evidence to interpret why literacy gaps exist and persist. In the findings step, that interpretation was translated into a set of strategic practices, each tied back to specific evidence. The method has clear limits. It is interpretive, not exhaustive; it does not weight studies statistically; and it reflects the author's selection choices. These limits are acceptable for a forward-looking conceptual article whose goal is to organise thinking rather than to estimate an effect size. 4. Analysis 4.1 Literacy gaps are structured, not random The first analytic point is that low genetic literacy is patterned. It is not spread evenly and then waiting for education to lift it uniformly. Evidence on general health literacy already shows that a large share of adults struggle to read and use health information; Kaphingst (2023) notes that limited literacy affects a substantial portion of the population and that genetic information adds further difficulty because it is abstract and probabilistic. When literacy is layered on top of unequal cultural capital, the people least able to interpret a #polygenic_risk_score are often the same people with the least access to follow-up care (Halbert, 2022; Paccoud et al., 2022). Read through Bourdieu, this is predictable. Genomic data enters a field already organised by capital. Those with the #habitus to engage confidently with clinicians extract more value from the same information. The Japanese study by Miyoshi and colleagues (2023) supports this: components of #genomic_literacy, including knowledge and numeracy, were linked to how people perceived risks and made decisions about #genetic_testing. People do not respond to genetic information in the same way; they respond according to the resources they bring to it. 4.2 The global gap mirrors the world-system The second analytic point scales the problem up. The same logic that sorts individuals within a country sorts countries within the world. Core nations hold the infrastructure and the data; peripheral nations hold the need but not the capacity (World Health Organization Science Council, 2022). Because most reference data comes from a narrow set of populations, the very tools meant to personalise care are least accurate for those outside the core. A health system cannot build genetic literacy in its population if it must rely on curricula, software, and professional norms produced elsewhere and not adapted to its languages, ancestries, and disease patterns. This produces a literacy version of unequal exchange. Peripheral populations may be studied, sampled, and surveyed, generating data and publications that accumulate value in the core, while the educational dividend, the trained counsellors and the locally relevant materials, does not flow back in equal measure. Future forecasting that assumes literacy will simply trickle outward from the core repeats the optimism that world systems theory warns against. Without deliberate redistribution of capacity, the global literacy gap is likely to widen as genomic medicine advances. 4.3 Imitation without adaptation The third analytic point concerns the remedies. Faced with the pressure to "do something" about genetic literacy, health systems often look sideways at what respected peers are doing and copy it. This mimetic behaviour is exactly what institutional isomorphism predicts under uncertainty (Powell & DiMaggio, 2023). Professional networks spread common #genetic_counseling models and common measurement instruments through normative channels, and funders attach conditions that push systems toward similar structures through coercive channels. The danger is that copied tools carry hidden assumptions. A literacy instrument validated in one language and reading level may misclassify understanding in another. A counselling model built around one-to-one specialist visits may be unworkable where there are few specialists. The systematic review by Daly and Kaphingst (2023) shows how unstable measurement already is even within similar settings; transplanting those measures across very different contexts compounds the problem. #Policy_diffusion can therefore create the appearance of progress, with adopted programmes and reported metrics, while leaving actual understanding untouched. Isomorphism explains why so many literacy efforts look alike and why so many disappoint. 4.4 Commercial pressure adds a fourth dynamic Beyond the three theories, the market deserves attention because it is reshaping the field faster than public systems can respond. Direct to consumer testing puts genetic results into people's hands with little or no professional support. Systematic reviews find that consumers often misunderstand results, that some seek confirmation or care from health services afterwards, and that the consequences for both individuals and systems are mixed (Nolan & Ormondroyd, 2023; Van Steijvoort et al., 2025). Policy frameworks for governing these services are still being developed (Onstwedder, Jansen, Cornel, & Rigter, 2024). The market does not wait for literacy to catch up; it sells first and assumes understanding later. This makes literacy not just an educational goal but a consumer-protection goal. 4.5 Trust, ancestry bias, and data ownership A further layer sits underneath all three theories: trust. Literacy is not only about understanding facts; it is about whether people believe the institutions delivering those facts have their interests at heart. Communities that have experienced discrimination or exploitation in research often approach genomic programmes with justified caution, and that caution is rational rather than ignorant. Paccoud and colleagues (2022) show that healthcare navigation is shaped by experiences of discrimination, not only by knowledge. A literacy strategy that ignores trust will read as persuasion rather than education, and it will fail in exactly the groups it most needs to reach. Trust is also tangled up with ancestry bias. Because most reference data comes from populations of European descent, risk tools frequently perform less well for everyone else, which means a person from an under-represented group may be handed a result that is both confusing and unreliable. This compounds the cultural capital problem: the people with the least support to interpret a result are often given the least accurate result to interpret. Finally, questions of data ownership matter for literacy in the long run. When peripheral populations contribute genomic material but have little say over how it is stored, used, or commercialised, the sense of shared ownership that sustains public understanding erodes. Building literacy therefore depends partly on building fair arrangements for whose data is used and who benefits, a point that connects everyday understanding back to the global structures described by world-systems theory. 5. Findings: Strategic Practices for the Next Decade The analysis points to a set of strategic practices. None is sufficient alone. Together they form a programme that addresses inequality within societies, imbalance between societies, and the risk of imitation without adaptation. 5.1 Standardise how literacy is defined and measured The field cannot improve what it cannot consistently measure. Daly and Kaphingst (2023) show that inconsistent definitions and instruments make studies hard to compare and progress hard to track. A strategic priority is to agree on a shared, layered definition of genetic literacy covering factual knowledge, numeracy, and critical literacy, and to validate a small set of measures across languages and reading levels. Standard measurement should be treated as shared public infrastructure rather than as something each project rebuilds, and crucially it should be re-validated locally rather than copied wholesale, which guards against the isomorphic trap. 5.2 Build a learning pathway from school to adulthood Literacy cannot be installed in a single clinic visit. The most durable gains come from teaching genetic concepts early and reinforcing them across life. This means embedding clear, accurate genetics in school #science_education, and then continuing through accessible adult and community learning. Because reading levels vary widely, materials should be written plainly; one review noted that average reading ability in some high-income populations sits at around a nine-year-old level, and that a meaningful share of adults have very limited literacy (Gasteiger et al., 2023). Designing for the least confident reader, not the average one, is the practical rule. 5.3 Expand and diversify the genomics workforce As specialists become scarce relative to demand, literacy work must spread to a wider workforce: primary care staff, nurses, pharmacists, community health workers, and trained peer educators. Expanding the workforce is also a way to break the core-periphery pattern, because locally trained professionals can teach in local languages and contexts. Investing in genetic counseling capacity in under-served regions is one of the clearest responses to the global imbalance that world systems theory describes (World Health Organization Science Council, 2022; Halbert, 2022). 5.4 Communicate in ways that match people's cultural capital If genetic knowledge behaves like cultural capital, then health communication must be designed to reach those with less of it, not only those who already engage easily. This means plain-language explanations, visual and numerical aids for risk, attention to trust, and respect for the cultural and family settings in which people make decisions. Luo and colleagues (2023) show that #social_capital and community networks carry health understanding, which suggests that community-based and family-oriented approaches will often outperform individual leaflets. The goal is to redistribute understanding, deliberately closing the gap that capital would otherwise widen. 5.5 Use digital tools, but check who they reach Digital education can scale quickly and personalise learning. A systematic review found that digital interventions can support genetics education, empowerment, and service engagement (Gasteiger et al., 2023). But digital delivery risks excluding those without devices, connectivity, or digital confidence, who are often the same groups with lower cultural capital. The strategic rule is to pair digital tools with non-digital options and to measure reach across social groups, so that technology narrows rather than widens the #health_disparities it is meant to address. 5.6 Regulate direct-to-consumer testing as a literacy issue Because the commercial market moves fastest, governing it is a literacy strategy in itself. Practical measures include clear standards for how results and uncertainty are presented, plain-language explanations of what a result does and does not mean, and routes to professional support when needed. Emerging policy frameworks for direct to consumer testing services offer a starting template (Onstwedder et al., 2024), and review evidence on consumer outcomes can guide where the protections matter most (Nolan & Ormondroyd, 2023; Van Steijvoort et al., 2025). 5.7 Rebalance global capacity through South-led partnership Finally, the global gap must be addressed directly. This means investing in laboratory infrastructure, training, and reference data in low- and middle-income regions, and ensuring that partnerships are led by, not merely located in, those regions. The aim is to move capacity toward the periphery rather than extracting data from it, so that populations can build genomic literacy using locally relevant tools and professionals (World Health Organization Science Council, 2022). Without this, future forecasting predicts a widening divide; with it, literacy can grow where need is greatest. 5.8 Summary of findings Read together, the practices form a coherent answer to the three-layer problem set out earlier. Standard measurement, life-course learning, workforce expansion, and capital-aware communication address inequality within societies. Workforce investment and South-led partnership address imbalance between societies. Local re-validation of tools and the pairing of digital with non-digital options address the risk of imitation without adaptation. The market-focused practice protects people during the period before literacy catches up. The common thread is that each practice treats genetic literacy as a structural and equity question, which is the central argument of this article. 6. Conclusion Genomic medicine is becoming a standard part of #public_healthcare faster than populations are gaining the knowledge to use it well. Kaphingst (2023) is right to frame this as a matter for future forecasting: the decisions made now about how to teach, measure, and govern genetic literacy will shape who benefits from genomics for a generation. The temptation is to treat literacy as a simple information problem, solved by producing more material and hoping people absorb it. This article has argued that such an approach will fail, and worse, will widen the gaps it claims to close. The three theories used here explain why. Bourdieu shows that genetic knowledge behaves like cultural capital, so that broadcasting information rewards the already-advantaged first. World systems theory shows that genomic capacity is concentrated in a wealthy core, leaving peripheral populations dependent and under-served. Institutional isomorphism shows that health systems tend to copy literacy models without adapting them, producing the look of progress without its substance. Each theory turns a seemingly individual deficit into a structured outcome of resources, hierarchy, and imitation. The strategic response must therefore work on all three levels at once. Within societies, that means standard but locally validated measurement, learning that runs from school to adulthood, a broader and more diverse workforce, and health communication built for those with the least cultural capital. Between societies, it means moving real capacity toward the periphery through South-led partnership rather than extracting data from it. Across the field, it means spreading good practice while resisting the unthinking policy diffusion that isomorphism encourages. And because the commercial market will not wait, regulating direct to consumer testing belongs inside the literacy strategy, not outside it. Several limitations and open questions should temper these conclusions. This article is an interpretive synthesis, not an empirical test, so its claims are arguments to be examined rather than findings to be applied directly. The evidence base is also uneven: much of it comes from high-income countries, which means the picture of peripheral need is drawn partly from the outside. Future research should prioritise three things. It should develop and validate shared literacy measures across languages and reading levels so that progress can be tracked fairly. It should test which strategic practices actually raise understanding in under-served groups, rather than assuming that programmes proven in the core will transfer. And it should study trust directly, since trust may explain more of the variation in how people use genetic information than knowledge alone. Until that work is done, forecasts in this field should be held with appropriate humility. Success would look like this: a population in which a person of modest means, in any region, can receive a genetic result, understand what it means for them and their family, ask informed questions, and make a decision that reflects their own values. That outcome is achievable, but only if genetic literacy is designed into genomic medicine as an #health_equity commitment from the beginning. The future of genomics will not be decided only in the laboratory. It will be decided by whether ordinary people are equipped to understand what the laboratory produces. #Patient_empowerment, in the end, is the real measure of whether genomic progress has been shared or simply announced. Hashtags #genetic_literacy #genomic_medicine #public_health #health_equity #cultural_capital #world_systems_theory #institutional_isomorphism #precision_medicine #direct_to_consumer_testing #polygenic_risk_scores #genetic_counseling #health_literacy #future_forecasting #patient_empowerment #public_health_genomics References Abel, T., & Benkert, R. (2022). Critical health literacy: Reflection and action for health. Health Promotion International, 37(4), daac114. https://doi.org/10.1093/heapro/daac114 Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241–258). Greenwood Press. Daly, B., & Kaphingst, K. A. (2023). Variability in conceptualizations and measurement of genetic literacy. PEC Innovation, 2, 100147. https://doi.org/10.1016/j.pecinn.2023.100147 DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101 Gasteiger, N., Vercell, A., Dowding, D., van der Veer, S., Brown, P., & Dunn Lopez, K. (2023). Digital interventions for genomics and genetics education, empowerment, and service engagement: A systematic review. Journal of Community Genetics, 14(3), 227–240. https://doi.org/10.1007/s12687-023-00648-w Halbert, C. H. (2022). Equity in genomic medicine. Annual Review of Genomics and Human Genetics, 23, 27–43. https://doi.org/10.1146/annurev-genom-112921-022635 Kaphingst, K. A. (2023). Future forecasting for research and practice in genetic literacy. Public Health Genomics, 26(1), 159–164. https://doi.org/10.1159/000533968 Luo, Y., Zhao, H., Chen, H., & Xiao, M. (2023). Association between cultural capital and health literacy during the COVID-19 pandemic among community residents in China: The mediating effect of social capital. Frontiers in Public Health, 11, 1199941. https://doi.org/10.3389/fpubh.2023.1199941 Miyoshi, T., Watanabe, M., & Ayub, Q. (2023). Impact of genomic literacy components on genetic testing decision-making in the general Japanese population in the 20s and 30s. PLOS ONE, 18(3), e0283432. https://doi.org/10.1371/journal.pone.0283432 Nolan, J. J., & Ormondroyd, E. (2023). Direct-to-consumer genetic tests providing health risk information: A systematic review of consequences for consumers and health services. Clinical Genetics, 104(1), 3–21. https://doi.org/10.1111/cge.14332 Onstwedder, S. M., Jansen, M. E., Cornel, M. C., & Rigter, T. (2024). Policy guidance for direct-to-consumer genetic testing services: Framework development study. Journal of Medical Internet Research, 26, e47389. https://doi.org/10.2196/47389 Paccoud, I., Nazroo, J., & Leist, A. K. (2022). Region of birth differences in healthcare navigation and optimisation: The interplay of racial discrimination and socioeconomic position. International Journal for Equity in Health, 21(1), 106. https://doi.org/10.1186/s12939-022-01709-1 Powell, W. W., & DiMaggio, P. J. (2023). The iron cage redux: Looking back and forward. Organization Theory, 4(4). https://doi.org/10.1177/26317877231221550 Van Steijvoort, E., Goossens, K., Demesure, K., Stanczak, A., Siermann, M., & Borry, P. (2025). Sharing behavior and health care utilization following direct-to-consumer genetic testing: A systematic review. Open Research Europe. https://doi.org/10.12688/openreseurope.19751.2 Wallerstein, I. (2004). World-systems analysis: An introduction. Duke University Press. World Health Organization Science Council. (2022). Accelerating access to genomics for global health. World Health Organization.

#Newborn_screening (NBS) is one of the most successful public health programs of the last sixty years, catching a small set of treatable conditions in the first days of life. New work is now testing whether #genomic_sequencing can widen that net to hundreds or even thousands of genetic conditions. Whether this shift succeeds will depend less on the laboratory and more on whether parents trust it and understand it. This article examines #parental_perspectives on expanded NBS through sequencing and the #educational_frameworks that public systems will need to put in place. Anchored on the interview study by Timmins et al. (2022) and supported by peer-reviewed literature published between 2021 and 2026, the analysis uses an interpretive thematic synthesis. Three sociological lenses are applied: Bourdieu's theory of #cultural_capital, the model of #institutional_isomorphism, and #world_systems_theory. The synthesis finds that parents are broadly open to expansion but place clear limits around conditions that are untreatable, late in onset, or low in penetrance; that they value both medical and personal forms of usefulness; and that they want education that arrives early, is multilingual, and is shaped by people like themselves. Programs around the world are converging on similar designs through imitation and shared professional norms, while access remains split along global lines of wealth. A tiered, co-designed educational model is proposed to support equitable #public_health integration. The article argues that #genomic_literacy should be treated as a public resource to be distributed, not a private asset that families are assumed to already hold. Keywords: newborn screening; genomic sequencing; parental attitudes; genetic literacy; health equity; cultural capital; institutional isomorphism; public health genomics 1. Introduction For decades, a few drops of blood taken from a baby's heel have been enough to screen for a short list of serious, treatable disorders. This #heel_prick test, repeated millions of times a year, is the backbone of #newborn_screening. Conditions such as phenylketonuria, congenital hypothyroidism, and severe combined immunodeficiency can be caught before symptoms appear, and early treatment can prevent death or lifelong disability. The logic is simple and powerful: find the condition while there is still time to act. #Genomic_sequencing now offers a way to read a baby's DNA directly, rather than measuring the chemical by-products of disease. Supporters argue that sequencing could expand screening from a few dozen conditions to many hundreds, identifying more children who might benefit from early care (Ziegler & Chung, 2024). Large studies are already underway. The GUARDIAN program in New York is offering sequencing to newborns alongside standard screening, the BabySeq Project has tested it within a clinical trial, and the United Kingdom is moving toward sequencing through its national Generation Study (Armstrong et al., 2022; Rankin et al., 2025; Ziegler & Chung, 2024). More than thirty similar efforts exist internationally (Minten et al., 2025). Yet the technology is the easy part. The harder question is whether ordinary families will accept #genomic_newborn_screening, and whether public systems can explain it well enough for #informed_consent to mean something. A test that screens for thousands of conditions is not the same as a test for a handful. It raises possibilities of uncertain results, conditions with no treatment, adult-onset risks discovered in infancy, and questions about who keeps the data and for how long. These are not technical problems. They are social ones, and they sit squarely with parents. This is where the work of Timmins et al. (2022) becomes central. In a study of English- and Spanish-speaking mothers in the United States, the authors found that parents were interested in expansion but had clear conditions, and that they wanted more education, delivered earlier, in language and formats that fit their lives. The study concluded that the success of expanded NBS depends on #parent_education that is accessible, curated by trusted experts, and built with parents who reflect the diversity of the population. This article takes that conclusion as its starting point. It asks two linked questions. First, what do diverse parents actually want and fear when sequencing enters #public_health screening? Second, what kind of #educational_framework would meet those needs in a fair way? To answer them, the article reads recent evidence through three sociological theories — Bourdieu, #institutional_isomorphism, and #world_systems_theory — that help explain why knowledge, trust, and access are spread so unevenly. The argument is that genomic literacy functions as a form of #cultural_capital, that screening programs are copying one another in predictable ways, and that the global rollout risks repeating an old pattern in which the wealthiest regions benefit first. 2. Background and Theoretical Framework 2.1 From metabolite screening to genomic screening Traditional NBS rests on the screening principles set out by Wilson and Jungner in 1968: screen only for conditions that are serious, detectable early, and treatable, using a test that is acceptable to the population. #Genomic_sequencing strains every part of that logic. Sequencing can detect risk for conditions long before any treatment exists, and it can return findings whose meaning is uncertain. The clean line between "treatable" and "not treatable" blurs. A second problem is that there is little agreement on what to screen for. When researchers compared the gene lists used by different programs, they found striking variation. One analysis of seven programs found that fewer than five percent of genes appeared on every list (Downie et al., 2024). A comparison of twenty-seven programs found that only seventy-four genes, about 1.7 percent, were shared by more than eighty percent of them (Minten et al., 2025). Other reviews confirm wide differences in how programs pick both genes and diseases (Betzler et al., 2024). The strongest predictor of whether a gene is included is whether it already sits on an established panel, which suggests that programs lean on existing authority rather than starting fresh (Minten et al., 2025). This pattern of #gene_selection is important for the theory that follows. A third issue is that expanded screening produces results that families must interpret and live with. The Early Check program in the United States and the Screen4Care project in Europe have both shown that returning results — even normal ones — is a communication task as much as a laboratory one. This is why #parental_perspectives and #genomic_literacy cannot be treated as add-ons. They are the program. 2.2 Bourdieu: genomic literacy as cultural capital Pierre Bourdieu argued that societies distribute not only money but also #cultural_capital — the knowledge, language, comfort, and confidence that let a person move easily through powerful institutions (Bourdieu, 1986). Cultural capital comes in forms: an embodied form (how at ease you feel reading a consent form or asking a doctor a question), an objectified form (the books and materials you have access to), and an institutionalized form (degrees and credentials). People who hold more of it find clinical settings familiar; people who hold less can find the same settings confusing or intimidating, even when the staff mean well. Applied to #genomic_newborn_screening, this idea is sharp. The ability to understand penetrance, carrier status, variants of uncertain significance, and the difference between risk and certainty is a kind of cultural capital. Families who already possess it can give #informed_consent in a real sense. Families who do not are asked to consent to something they have had no chance to understand. Bourdieu also described #habitus — the deep, half-conscious dispositions that shape how people react to new situations. A parent whose habitus includes trust in medical institutions will respond very differently from one whose family history includes being experimented on or ignored. The study by Timmins et al. (2022) captured exactly this when it found that information needs differ by language and culture. The clinic, in Bourdieu's terms, is a #field — a structured space with its own rules, where doctors and genetic counselors hold #symbolic_power. When an expert frames sequencing as obviously beneficial, that framing carries weight that a parent may feel unable to challenge. An #educational_framework built on this insight would not simply hand out leaflets. It would try to redistribute cultural capital, lowering the barrier so that parents from every background can take part as something closer to equals. 2.3 Institutional isomorphism: why programs look alike Why do screening programs around the world, designed by different governments, end up resembling one another? DiMaggio and Powell's concept of #institutional_isomorphism offers an answer (DiMaggio & Powell, 1983). They argued that organizations facing the same environment tend to become similar through three pressures. The first is coercive isomorphism, where law and funding force a shape. In the United States, the Recommended Uniform Screening Panel acts as a coercive anchor, which is why being on an existing panel so strongly predicts inclusion in newer programs (Minten et al., 2025). The second is mimetic isomorphism, where organizations facing uncertainty copy others that seem successful. Because no one is sure which genes belong in #genomic_screening, new programs imitate established ones, which explains both the convergence on a small core of genes and the messy variation around the edges (Downie et al., 2024). The third is normative isomorphism, spread by shared professional training. Genetic counselors and clinical geneticists move between programs and carry common standards with them, pushing toward similar consent processes and similar education models (Coleman et al., 2023). This lens matters for education because it predicts that #educational_frameworks will also converge. If one program designs strong multilingual materials and is praised for it, others will copy the approach. That can spread good practice quickly. It can also spread weak practice quickly if a flawed model becomes the template everyone imitates. A program that treats parent education as a checkbox may pass that habit on to the next program through the same channels. Designing the early models well is therefore not only a local choice but a structural one. 2.4 World-systems theory: the global pattern of access Immanuel Wallerstein's #world_systems_theory describes a global economy split into a wealthy core, a dependent periphery, and a semi-periphery in between (Wallerstein, 2004). The core captures the highest-value activity; the periphery supplies inputs and receives the leftovers. Although Wallerstein wrote about trade, the pattern fits #genomic_medicine well. The most advanced #genomic_newborn_screening pilots are running in the United States, the United Kingdom, Australia, and parts of Western Europe (Rankin et al., 2025; White et al., 2023). Meanwhile, much of the world still lacks basic newborn screening at all. Commentators estimate that around two-thirds of the global population receives little or no NBS, and that in parts of Latin America large shares of newborns are screened for nothing (Tariq et al., 2025). The result is a widening gap: the core debates whether to screen for thousands of conditions while the periphery cannot yet screen for two. There is a deeper concern. Genomic databases built mostly from European-ancestry populations interpret variants less accurately for everyone else, which means the science itself can carry a core bias into every clinic that uses it (Tariq et al., 2025). If #health_equity is the goal, world-systems theory warns that simply exporting core technology will not close the gap and may even widen it. The flow of genomic data from periphery to core, without matching benefit returning, would repeat the very pattern Wallerstein described. Together these three lenses frame the rest of the article. Bourdieu explains the inequality inside the clinic, #institutional_isomorphism explains why programs and their education models copy each other, and #world_systems_theory explains the inequality between nations. 3. Method This article is an interpretive thematic synthesis rather than a systematic review or a report of new data. The aim was to draw together what is known about #parental_perspectives and #educational_frameworks for expanded NBS, and to interpret it through the three theories described above. The approach has four parts. First, a focal case was chosen. The interview study by Timmins et al. (2022) was selected because it speaks directly to both questions in this article — what diverse parents want, and what education they need — and because its participants included Spanish-speaking mothers whose views are often missing from the record. This study anchors the synthesis. Second, supporting literature was gathered. Searches in major biomedical and social science databases used combinations of terms such as newborn screening, genomic sequencing, parental attitudes, genetic literacy, education, consent, and equity. Priority went to peer-reviewed articles published between 2021 and 2026 so that the evidence would reflect the current state of the field. Foundational sociological texts were included where needed to define the #theoretical_framework. Items were kept if they reported empirical findings on parent or public attitudes, examined education or consent for #genomic_screening, or described program design and gene selection. Third, the material was coded thematically. Recurring ideas were grouped into themes through repeated reading. Five themes emerged: conditional support for expansion; medical versus personal usefulness; educational timing and format; trust, privacy, and data; and equity across language, culture, and nation. This stage was inductive, staying close to what parents actually said. Fourth, the themes were interpreted abductively against theory. Each theme was examined for what Bourdieu, #institutional_isomorphism, or #world_systems_theory could explain about it. This is where the synthesis moves from description to interpretation. The limits of this method should be stated plainly. A narrative synthesis reflects the judgment of its author and cannot claim the completeness of a registered systematic review. Much of the underlying evidence comes from high-income, English-speaking settings, which narrows the picture. No new parents were interviewed, so the article cannot speak for groups the existing studies missed. These limits are taken up again in the conclusion. 4. Analysis 4.1 Conditional support: parents say yes, with lines drawn Across the literature, the headline is that parents are interested but not uncritical. In the focal study, every participant supported some degree of expansion, yet they disagreed about which conditions belonged in it (Timmins et al., 2022). Support dropped for conditions with low penetrance, for those without an approved treatment, and for those whose effects appear only later in life. The BabySeq work found a similar mix of enthusiasm and caution, with many parents wanting sequencing offered to all newborns but hesitating over whether it should be mandatory or whether consent should be required (Armstrong et al., 2022). When asked specifically about adult-onset findings in their children, parents in the BabySeq Project were divided, weighing the value of warning against the burden of knowing too early (Pereira et al., 2023). A scoping review of public and parent views reached the same place from a wider angle: people are positive about #genomic_sequencing in NBS in principle, yet more willing to take part in standard screening than in genomic screening when asked to choose (drawing on the synthesis reported across recent reviews). Parents from diverse backgrounds, studied directly, also showed careful, conditional preferences rather than blanket approval (Gold et al., 2025). Read through Bourdieu, these conditions are not random. The lines parents draw — treatable, early, certain — mirror the original screening principles, which suggests that parents are reasoning with the same logic clinicians use, when they are given the chance. The gap is not in parents' judgment but in their access to the information that lets them apply it. Where #cultural_capital is high, conditional support is articulate and specific; where it is low, the same caution may appear as silence or refusal, which is easy to misread as disinterest. 4.2 Medical and personal usefulness A consistent finding is that parents value sequencing for more than medicine. In the focal study, parents endorsed clear medical benefits but also pointed to personal or non-medical value: the chance to prepare, to avoid being blindsided, to plan future pregnancies, and to make sense of family history (Timmins et al., 2022). This idea of #personal_utility recurs across studies of parental attitudes and decision-making (Pereira et al., 2023). This matters for design because a program built only around medical action will undersell what parents actually want and may also overpromise. If parents value preparedness, then education must be honest about uncertainty — about results that signal risk without certainty, and about conditions for which preparation, not cure, is all that medicine can offer. The concept of #genomic_literacy here is not just understanding genes; it is understanding what a result can and cannot do. 4.3 Educational timing and format The strongest practical message in the focal study is about timing and form. Parents wanted education earlier — before birth, even before pregnancy — rather than in the rushed hours after delivery (Timmins et al., 2022). They wanted information that is easy to reach, curated by trusted experts, and shaped by parents who look and sound like them. They also signaled that needs differ by language and culture, so a single English-language pamphlet will not do. Wider research on #genetic_literacy supports this. Reviews note that the field still measures literacy inconsistently and that interventions often assume more baseline knowledge than people have (Daly & Kaphingst, 2023; Kaphingst et al., 2023). Studies of decision-making show that literacy has several parts — factual knowledge, numeracy, and the interactive and critical skills needed to use information — and that all three shape how people weigh risks and benefits (Miyoshi et al., 2023). Community-based work shows that co-designed, multilingual, culturally responsive materials build trust and understanding better than generic ones, especially in groups that have reason to be wary. The #institutional_isomorphism lens adds a warning and an opportunity. Because programs copy each other, the timing and format chosen by early programs may become the default everywhere. If leading programs move education upstream into prenatal and preconception care and build it with communities, that good practice can spread through the same imitation and professional networks. If they bolt education onto the delivery room as an afterthought, that weakness will spread just as fast. 4.4 Trust, privacy, and data Expanded screening generates genomic data that lasts a lifetime, and parents notice. Concerns about privacy, storage, secondary use, and who can access a child's genome appear throughout the literature (Cao et al., 2023; Dikow et al., 2022). Healthcare professionals share these worries, raising practical and ethical questions about consent, data governance, and the risk of overdiagnosis (Cao et al., 2023). Trust is not evenly distributed, and here Bourdieu and history meet. Families from groups that have been mistreated by medical and research institutions bring a habitus shaped by that experience. For them, reassurance about data is not a formality; it is the whole question. A consent process that ignores this will read as #symbolic_power being exercised over them rather than information being shared with them. Building trust therefore means concrete commitments about data — clear limits on use, real choices about storage, plain answers about access — not just friendly language. 4.5 Equity across language, culture, and nation The final theme runs through all the others. Within countries, parents' needs differ by language and culture, and programs that ignore this will reach some families well and others poorly (Gold et al., 2025; Timmins et al., 2022). Between countries, the divide is starker. The pilots that define the cutting edge sit in a handful of wealthy nations, while most of the world lacks even basic NBS (Tariq et al., 2025; Rankin et al., 2025). This is where #world_systems_theory earns its place. Without deliberate effort, #genomic_newborn_screening will follow the path of other medical technologies: developed in the core, optimized for core populations, and reaching the periphery late, in diluted form, and interpreted through reference databases that fit core ancestries best. The danger is not only that the periphery waits longer. It is that the data and benefits flow toward the core while the gaps in care stay where they are. #Health_equity, on this reading, is not a side effect that good technology will produce on its own. It has to be designed in, at both the national and the global level. 5. Findings Bringing the analysis together produces a set of clear findings and a proposed framework. Finding 1: Parental support is real but conditional, and the conditions are reasonable. Parents back expansion while drawing lines around untreatable, late-onset, and uncertain conditions. These lines track the long-standing logic of screening, which means parents are partners in reasoning, not obstacles to it. Programs should treat conditional support as informed support and design choice into the system rather than assuming either blanket consent or blanket refusal. Finding 2: Parents value preparation as much as cure. The pull of #personal_utility means education must explain what results can and cannot offer, including honest handling of uncertainty. Overselling medical action while ignoring preparedness misreads what families want. Finding 3: Education must move upstream and be co-designed. The clearest practical demand is for earlier, accessible, multilingual information shaped by trusted experts and by parents from the communities served. This is a redistribution of #cultural_capital, not a marketing task. Finding 4: Trust depends on concrete data commitments. Privacy and data governance are central, not peripheral, especially for families with histories of medical mistreatment. Vague reassurance will not build the trust that #informed_consent requires. Finding 5: Equity must be engineered at two scales. Inequity within countries (language, culture, literacy) and between countries (core and periphery) will both widen by default. Closing them requires deliberate investment, not the hope that technology will spread fairly on its own. 5.1 A proposed tiered educational framework From these findings, a four-tier #educational_framework follows. It is designed to redistribute genomic literacy across the population rather than assuming families already hold it. The first tier is preconception and general public awareness. Basic, plain-language ideas about #newborn_screening and #genomic_sequencing should reach people before pregnancy, through schools, primary care, and community organizations, so that the topic is not entirely new when a baby arrives. This tier treats #genomic_literacy as a public good built over time. The second tier is prenatal engagement. During pregnancy, when there is time to think and ask questions, parents should receive deeper, interactive information about what expanded screening can find, what the limits are, and what choices they will face. Materials should be multilingual and co-designed with parents from the relevant communities, following the evidence that culturally responsive resources work better. The third tier is point-of-care consent. At or near birth, the focus shifts to a genuine consent conversation, supported by trained staff, that builds on the earlier tiers rather than starting from zero. Because the earlier tiers have done the heavy lifting, this conversation can be shorter, calmer, and more meaningful. The fourth tier is post-result support. Whatever the result, families need help interpreting it, including access to genetic counseling for complex or uncertain findings and clear, honest communication when results are normal. The Early Check experience shows that even normal results need careful explanation. Two cross-cutting supports hold the tiers together. One is #workforce_training: most births are attended by professionals who are not genetics specialists, so non-genetics clinicians need practical genomic education (Coleman et al., 2023). The other is data governance that is transparent and gives families real choices, which is the foundation of trust. 5.2 How theory shapes the framework Each tier answers a theoretical problem. The preconception and prenatal tiers attack the #cultural_capital gap that Bourdieu identifies, moving knowledge to people before the high-pressure moment when inequality bites hardest. The emphasis on co-design and multilingual materials is a direct attempt to redistribute capital rather than reward those who already have it. #Institutional_isomorphism tells us that whatever early programs build will be copied. That makes the case for getting the framework right now, because a strong model will spread through the same imitation and professional networks that currently spread weak ones. Shared professional standards for #workforce_training are a way to turn normative isomorphism into a force for quality. #World_systems_theory shapes the framework's ambition. A model designed only for wealthy, English-speaking, high-literacy populations will deepen the core-periphery divide. Building multilingual, low-literacy-friendly, community-led education from the start, and sharing it openly, is a way to resist that pull. At the global level, the same logic argues for investing in basic screening capacity in the periphery rather than exporting only the most advanced and most data-hungry tools. 6. Conclusion Reading a newborn's genome is becoming technically routine. Making it socially fair is not. The evidence is consistent: parents are willing to expand #newborn_screening through #genomic_sequencing, but they want a real say, honest information about uncertainty, strong protection for their child's data, and education that arrives early and speaks their language. The study by Timmins et al. (2022) said this plainly, and the wider literature has confirmed it across countries and methods. The three theories used here turn these findings into a coherent argument. Bourdieu shows that #genomic_literacy is a form of #cultural_capital that families hold unequally, so education must redistribute it rather than assume it. #Institutional_isomorphism shows that screening programs and their education models copy one another, which means the designs chosen now will echo widely and should be built with care. #World_systems_theory shows that, left alone, expanded screening will favor the wealthy core and may widen the global gap in care. The proposed tiered, co-designed #educational_framework is one attempt to answer all three problems at once. This article has limits. It is an interpretive synthesis, not a systematic review, and it leans on evidence drawn mainly from high-income, English-speaking settings. The voices most likely to be missing — families in low-resource regions, speakers of less-studied languages, communities with the deepest reasons for mistrust — are exactly the ones whose perspectives matter most for #health_equity. Future research should gather #parental_perspectives directly from these groups, test multilingual and co-designed educational models in real programs, and study whether the tiers proposed here actually raise genuine understanding rather than mere confidence. Work is also needed on data governance arrangements that parents trust and on building basic screening capacity where none yet exists. The central claim is simple. The success of #genomic_newborn_screening will be decided less by sequencing machines than by whether #public_health systems treat parents as partners and treat knowledge as something to be shared widely and fairly. If genomic literacy is left as a private asset, expanded screening will help those who already have the most. If it is built as a public resource, the promise of catching disease early could finally reach everyone. Hashtags #Newborn_Screening #Genomic_Sequencing #Parental_Perspectives #Public_Health_Genomics #Genomic_Literacy #Health_Equity #Cultural_Capital #Institutional_Isomorphism #World_Systems_Theory #Informed_Consent #Genetic_Counseling #Precision_Public_Health #Newborn_Genomics #Early_Detection #Parent_Education References Armstrong, B., Christensen, K. D., Genetti, C. A., Parad, R. B., Robinson, J. O., Blout Zawatsky, C. L., Zettler, B., Beggs, A. H., Holm, I. A., Green, R. C., McGuire, A. L., Smith, H. S., & Pereira, S. (2022). Parental attitudes toward standard newborn screening and newborn genomic sequencing: Findings from the BabySeq Study. Frontiers in Genetics, 13, 867371. https://doi.org/10.3389/fgene.2022.867371 Betzler, I. R., Hempel, M., Mütze, U., Kölker, S., Winkler, E., Dikow, N., et al. (2024). Comparative analysis of gene and disease selection in genomic newborn screening studies. Journal of Inherited Metabolic Disease, 47(5), 945–970. Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241–258). Greenwood Press. Cao, M., Notini, L., Ayres, S., & Vears, D. F. (2023). Australian healthcare professionals' perspectives on the ethical and practical issues associated with genomic newborn screening. Journal of Genetic Counseling, 32(2), 376–386. Coleman, T., Bensend, T., Mills, R., Orlando, L. A., & Doyle, L. (2023). Critical components of genomic medicine practice for non-genetics healthcare professionals: Genetic counselors' perspectives and implications for medical education. Journal of Genetic Counseling, 32(4), 798–811. https://doi.org/10.1002/jgc4.1689 Daly, B., & Kaphingst, K. A. (2023). Variability in conceptualizations and measurement of genetic literacy. PEC Innovation, 2, 100147. DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. Dikow, N., Ditzen, B., Kölker, S., Hoffmann, G. F., & Schaaf, C. P. (2022). From newborn screening to genomic medicine: Challenges and suggestions on how to incorporate genomic newborn screening in public health programs. Medizinische Genetik, 34(1), 13–20. Downie, L., Bouffler, S. E., Amor, D. J., Christodoulou, J., Yeung, A., Horton, A. E., et al. (2024). Gene selection for genomic newborn screening: Moving toward consensus? Genetics in Medicine, 26(2), 101077. Gold, N. B., Omorodion, J. O., Del Rosario, M. C., et al. (2025). Preferences of parents from diverse backgrounds on genomic screening of apparently healthy newborns. Journal of Genetic Counseling, 34(2), e1994. Kaphingst, K. A., Daly, B., & Hovick, S. R. (2023). Future forecasting for research and practice in genetic literacy. Public Health Genomics, 26(1), 159–164. https://doi.org/10.1159/000533968 Minten, T., Gold, N. B., Bick, S., Adelson, S., Gehlenborg, N., Amendola, L. M., et al. (2025). Data-driven consideration of genetic disorders for global genomic newborn screening programs. Genetics in Medicine, 27(8), 101443. https://doi.org/10.1016/j.gim.2025.101443 Miyoshi, T., Watanabe, M., & Ayub, Q. (2023). Impact of genomic literacy components on genetic testing decision-making in the general Japanese population in the 20s and 30s. PLoS ONE, 18(3), e0283432. https://doi.org/10.1371/journal.pone.0283432 Pereira, S., Gutierrez, A. M., Robinson, J. O., Christensen, K. D., Genetti, C. A., Blout Zawatsky, C. L., et al. (2023). Parents' decision-making regarding whether to receive adult-onset only genetic findings for their children: Findings from the BabySeq Project. Genetics in Medicine, 25(1), 100002. Rankin, S., Jones, R., Boardman, F., Lucassen, A., & Hall, A. (2025). The UK National Screening Committee, the newborn genomes programme, and the ethical conundrum for UK newborn screening. Journal of Community Genetics. https://doi.org/10.1007/s12687-025-00788-1 Tariq, H., Salpietro, V., Houlden, H., & Stevanin, G. (2025). Globalizing newborn screening: Bridging gaps in genetic diagnosis and treatment. Frontiers in Pediatrics, 13, 1672993. https://doi.org/10.3389/fped.2025.1672993 Timmins, G. T., Wynn, J., Saami, A. M., Espinal, A., & Chung, W. K. (2022). Diverse parental perspectives of the social and educational needs for expanding newborn screening through genomic sequencing. Public Health Genomics, 25(5–6), 185–192. https://doi.org/10.1159/000526382 Wallerstein, I. (2004). World-systems analysis: An introduction. Duke University Press. White, S., Mossfield, T., Fleming, J., Barlow-Stewart, K., Ghedia, S., Dickson, R., et al. (2023). Expanding the Australian Newborn Blood Spot Screening Program using genomic sequencing: Do we want it and are we ready? European Journal of Human Genetics, 31(6), 703–711. Ziegler, A., & Chung, W. K. (2024). Universal newborn screening using genome sequencing: Early experience from the GUARDIAN study. Pediatric Research. https://doi.org/10.1038/s41390-024-03647-w

The emergence of #generative_AI, particularly #large_language_models (LLMs), is reshaping the landscape of #personalized_oncology and #clinical_decision_support. Grounded in the foundational diagnostic study by Benary et al. (2023), this article examines how LLMs are being integrated into the #precision_medicine workflow, enabling oncologists to navigate an increasingly complex terrain of #molecular_biomarkers, treatment guidelines, and patient-specific data. Drawing on Pierre Bourdieu's concepts of field and capital, world-systems theory, and the sociological framework of #institutional_isomorphism, this paper argues that the adoption of #AI_in_healthcare is not merely a technical transition but a deeply structured sociotechnical process shaped by institutional pressures, professional hierarchies, and global inequalities in access to advanced medical infrastructure. Through a narrative synthesis of recent empirical studies, this article analyses the performance, promise, and limitations of LLMs as decision-support tools in #tumor_board settings, #genomic_data interpretation, and personalized treatment matching. Findings suggest that while LLMs do not yet match the quality of human expert recommendations, they augment clinical reasoning, accelerate literature synthesis, and hold particular promise when deployed within retrieval-augmented generation (RAG) pipelines. Key barriers to adoption — including #AI_hallucination, data governance, equity of access, and the absence of standardized evaluation benchmarks — are discussed in depth. The article concludes with recommendations for responsible integration of #AI_tools in oncology practice. Keywords: generative artificial intelligence, large language models, oncology, clinical decision support, precision medicine, institutional isomorphism, Bourdieu, personalized treatment 1. Introduction Cancer remains one of the most diagnostically and therapeutically complex domains in modern medicine. Each patient presents a unique constellation of #tumor_genetics, co-morbidities, prior treatment histories, and psychosocial factors that resist simple, protocol-driven management. The rise of #precision_oncology over the past two decades has only deepened this complexity: oncologists must now interpret vast volumes of #next_generation_sequencing data, cross-reference rapidly evolving clinical guidelines, and identify off-label or trial-based options that may benefit individual patients — all within the compressed timescales of clinical consultations. Into this complexity, a new class of computational tool has arrived. #Large_language_models — artificial intelligence systems trained on enormous corpora of text and capable of generating coherent, contextually appropriate responses — have begun to attract serious attention from the oncology community. These systems, ranging from OpenAI's GPT-4 to open-source alternatives like Llama 2, can read, summarise, and reason across large bodies of biomedical literature, offering a potentially scalable path toward real-time #clinical_decision_support. The pivotal study underpinning this article is that of Benary et al. (2023), published in JAMA Network Open. That study evaluated four LLMs — ChatGPT, Galactica, Perplexity, and BioMedLM — using ten fictional patients with advanced cancers and genetic alterations. Each case was submitted to both the LLMs and a human expert, with #treatment_options then reviewed by a molecular #tumor_board. The study found that while LLMs deviated substantially from expert recommendations, they nonetheless identified several important treatment strategies and produced suggestions that were genuinely useful to the board. Crucially, at least one useful treatment option per patient was generated only by an LLM. This finding captures the essential tension at the heart of #AI_in_oncology: these tools are not ready to replace human clinical judgement, but they are already capable of augmenting it in meaningful ways. This article examines that tension through multiple lenses — technical, institutional, sociological, and ethical — to assess where #AI_decision_support is heading and what conditions are necessary for its responsible integration into clinical care. The article proceeds as follows. Section 2 establishes the theoretical framework, drawing on Bourdieu's field theory, world-systems analysis, and institutional isomorphism. Section 3 describes the methodological approach used in this narrative review. Section 4 analyses the current state of the evidence. Section 5 presents the key findings. Section 6 concludes with implications for practice and policy. 2. Background and Theoretical Framework 2.1 The Clinical Decision-Making Field in Oncology Pierre Bourdieu's concept of the field — a structured social space in which agents compete for forms of capital — offers a productive lens through which to understand #AI_integration in oncology. The oncological field is organised around multiple forms of capital: the scientific capital of those who produce cutting-edge research, the medical capital held by experienced clinicians at elite academic centres, and the institutional capital of hospitals and tumour boards that confer legitimacy on treatment decisions. The introduction of #AI_tools into this field does not occur in a neutral space; it disrupts existing hierarchies, challenges professional authority, and reconfigures what counts as valid knowledge. For Bourdieu, agents in a field engage in practices shaped by their habitus — the durable dispositions formed through prior socialisation. Senior oncologists trained in an era of purely expert-driven decision-making may resist the integration of AI not because they are unreasonable, but because such integration is at odds with their professional habitus. Conversely, junior clinicians and informaticists, whose habitus has been shaped by a digitally saturated training environment, may embrace these tools more readily. The uneven uptake of #AI_decision_support systems across different clinical settings and career stages reflects precisely this Bourdieuisian dynamic. 2.2 World-Systems Theory and Global AI Inequity Immanuel Wallerstein's world-systems theory draws attention to the structural inequalities between core, semi-periphery, and periphery nations in the global economy. Applied to #AI_in_healthcare, this framework reveals a troubling geography: the most sophisticated LLM-based #clinical_decision_support tools are being developed, validated, and deployed predominantly in high-income, technologically advanced countries — the United States, Germany, Japan, and the United Kingdom. The datasets on which these models are trained are largely drawn from patient populations in these same nations, meaning that their accuracy and clinical relevance may be systematically lower for patients in low- and middle-income countries where cancer burden is disproportionately high. From a world-systems perspective, the global diffusion of AI in oncology risks replicating — rather than reversing — structural inequities in #cancer_care. Peripheral nations that lack the computational infrastructure, curated clinical datasets, and regulatory capacity to develop their own systems may become dependent on imported tools that do not adequately reflect their patient populations, disease epidemiology, or healthcare contexts. This concern is not merely theoretical. As Benson et al. (2024) noted in their narrative review, applied LLM methods in cancer care still lag significantly behind advances in computer science, and standardised evaluation metrics are absent across the field. The gap between technological possibility and equitable deployment is a structural feature of how AI in medicine is currently organised. 2.3 Institutional Isomorphism and the Spread of AI Tools DiMaggio and Powell's concept of institutional isomorphism describes the process by which organisations come to resemble one another not because of efficiency gains but because of pressures to appear legitimate within a shared institutional field. Three mechanisms drive this convergence: coercive isomorphism (pressure from regulators and funders), mimetic isomorphism (copying what successful peers are doing), and normative isomorphism (adoption of practices endorsed by professional associations). All three mechanisms are visible in the spread of #AI_decision_support tools across oncology. Regulatory signals — such as the U.S. FDA's growing framework for AI-enabled medical devices and the European Union AI Act — exert coercive pressure. Institutional competition between academic cancer centres creates mimetic pressure: when Memorial Sloan Kettering Cancer Center or Charité Berlin integrates LLM-based tools, peer institutions face pressure to follow suit. Professional societies, through clinical guidelines and position statements, exert normative pressure. The result is a wave of adoption that may outrun clinical evidence, as institutions race to signal modernity rather than waiting for robust #randomized_clinical_trial validation. Currie's (2012) longitudinal study of electronic health records adoption in the UK National Health Service demonstrated precisely this pattern, showing that institutional isomorphic pressures led to adoption ahead of proven benefit, resulting in resistance and delayed implementation. There is a genuine risk that #LLM_integration in oncology follows a similar trajectory. 3. Method This article employs a narrative review methodology, synthesising findings from peer-reviewed empirical studies, systematic reviews, and high-quality preprints published between 2021 and 2026. The focus is on studies examining the application of LLMs and #generative_AI to clinical decision support in oncology, with particular attention to studies that evaluate model performance against expert or guideline standards. The search terms used to identify relevant literature included: "large language models oncology," "generative AI clinical decision support cancer," "LLM precision medicine," "GPT tumour board," and related variants. Priority was given to studies published in Scopus-indexed Q1 and Q2 journals, including JAMA Network Open, Nature Cancer, BMJ Oncology, JCO Clinical Cancer Informatics, Annals of Oncology, and Artificial Intelligence Review. Studies were selected if they included empirical evaluation of LLM performance, quantitative benchmarking against human expert decisions, or systematic analysis of LLM applications across cancer types. The theoretical framework sections draw on classical and contemporary works in sociology, institutional theory, and science and technology studies to contextualise findings within broader sociotechnical dynamics. Sociological theories were applied interpretively rather than formally coded, following established traditions of theoretically informed narrative review in medical sociology and health informatics. No primary data were collected for this study. Ethical approval was not required. 4. Analysis 4.1 LLMs as Decision-Support Tools: Performance Evidence The most methodologically rigorous evidence base for LLM performance in #clinical_decision_support comes from studies that directly compare model outputs with decisions made by expert oncologists or multidisciplinary tumour boards. The Benary et al. (2023) study established the benchmark. Across ten complex patient cases, the four LLMs achieved median F1 scores of 0.04, 0.17, 0.14, and 0.19 when compared against a human expert as the gold standard. When the outputs of all LLMs were combined, precision and recall each reached 0.29, suggesting that ensemble use of models offers modest improvement. Crucially, the molecular tumour board judged at least one LLM-generated treatment option per patient as useful, including novel strategies not identified by the human annotator. Subsequent research has built on and extended this baseline. Ferber et al. (2025), publishing in Nature Cancer, developed an autonomous AI agent integrating GPT-4 with precision oncology tools including vision transformers for microsatellite instability detection and web-based search capabilities. Evaluated on twenty realistic multimodal patient cases, the integrated system correctly used appropriate tools in 87.5% of cases, reached correct clinical conclusions in 91% of instances, and accurately cited oncology guidelines 75.5% of the time. Notably, GPT-4 alone achieved only 30.3% accuracy on the same cases, demonstrating that isolated LLM capability is insufficient and that performance rises dramatically when models are embedded within structured tool ecosystems. Studies examining LLM concordance with multidisciplinary tumour board decisions have largely found moderate-to-high agreement in well-defined clinical scenarios. Erdat et al. (2025), in a cross-sectional study published in Clinical and Translational Oncology, evaluated GPT-4 against 610 tumour board decisions at Ankara University Hospitals. The mean compatibility score between GPT-4 predictions and MTB decisions was 3.59 on a 4-point scale, with excellent inter-rater reliability. Discrepancies clustered around rare cancers and complex comorbid presentations, areas where current guidelines offer limited guidance and where the model's training data may be sparse. Büyükceran et al. (2025) reported similarly encouraging results in breast cancer specifically, finding that GPT-4o achieved full concordance with MTB decisions in 31 of 33 cases (93.9%) when provided with structured clinical input. The mean concordance score was 4.94 out of 5.0, with perfect inter-rater reliability. The two discordant cases involved questions about genetic counselling referral, suggesting that the model's limitations are not primarily in treatment selection but in the more nuanced, patient-centred dimensions of care planning. Less encouraging results emerged from the breast cancer benchmarking study by Shah et al. (2025/2026), a multi-institutional expert evaluation comparing three LLMs — GPT-4o, DeepSeek-R1, and OpenEvidence — across ten complex breast cancer scenarios rated by sixteen breast medical oncologists. OpenEvidence achieved the highest overall performance score of 3.91 out of 5, significantly outperforming both GPT-4o (3.19) and DeepSeek-R1 (2.93). GPT-4o showed the highest variability in expert ratings, reflecting inconsistent performance across cases. This variability is not a trivial concern: in clinical practice, an unpredictable tool is in some ways more dangerous than a consistently limited one, because clinicians cannot reliably calibrate when to override its suggestions. 4.2 Retrieval-Augmented Generation and Precision Oncology A consistent theme emerging from the evidence base is that #retrieval_augmented_generation (RAG) represents the most promising near-term pathway for LLM deployment in #precision_oncology. In a RAG pipeline, the LLM's responses are grounded in dynamically retrieved, trusted external knowledge sources — such as curated oncology databases, clinical guidelines, and peer-reviewed literature — rather than relying solely on patterns encoded in its weights during training. Jun et al. (2026), publishing in Cancer Cell, developed a RAG-LLM workflow using the Molecular Oncology Almanac (MOAlmanac) as a knowledge base and benchmarked it against an LLM-only approach. The RAG pipeline achieved up to 95% accuracy on synthetic queries and 93% accuracy on real-world queries from practising oncologists. By contrast, the LLM-only approach performed substantially worse. The authors concluded that this system could serve as actionable guidance for deploying LLMs in precision oncology clinical settings. Similar findings were reported by Kreimeyer et al. (2024), who demonstrated that a RAG pipeline using trusted treatment relationships from OncoKB reproduced over 80% of those relationships when queried using simple questions posed to an untrained Llama 2 model. This finding is significant because it demonstrates that even a general-purpose model can achieve high accuracy in oncological treatment matching if provided with reliable, structured context at inference time — without expensive fine-tuning on proprietary clinical datasets. For Hamamoto et al. (2026), in a comprehensive review published in the Journal of Hematology and Oncology, a RAG strategy combined with human-in-the-loop (HITL) workflows represents the most viable near-term model for responsible integration of #generative_AI in precision oncology. The authors describe a governance architecture centred on standardised clinical data models (OMOP, mCODE, FHIR), multicenter prospective evaluation, auditable decision logs, and synthetic data generation with differential privacy protections. This architecture is explicitly designed to address the #AI_governance challenges that have slowed clinical adoption. 4.3 Multimodal Applications and Tumour Board Support Cancer diagnosis and management is inherently multimodal, relying on integrated interpretation of pathology reports, radiology images, genomic sequencing outputs, clinical notes, and laboratory results. Early LLM applications in oncology were limited to text-based inputs, but the evidence base is rapidly expanding to include multimodal foundation models capable of processing images alongside text. Vasanthapuram (2025) described an AI pipeline combining generative models with LLMs for early cancer detection from radiological imaging, demonstrating improved diagnostic accuracy and reduced report generation time. Ferber et al. (2025) integrated vision transformers capable of detecting microsatellite instability, KRAS, and BRAF mutations from histopathology slides with language model reasoning about treatment implications, demonstrating that the combination substantially outperforms text-based reasoning alone. In tumour board settings specifically, Wang et al. (2025), publishing in iScience, synthesised evidence across cancer types and found that AI recommendations agreed with MDT decisions in 70–90% of cases, with the highest concordance observed for treatment intent, surgical decisions, and radiotherapy planning. Systemic therapy decisions showed the lowest AI-human agreement, reflecting the greater complexity and nuance of chemotherapy and immunotherapy selection. Nardone et al. (2024), reviewing AI's role in #multidisciplinary_tumor_boards from the perspectives of surgeons, medical oncologists, and radiation oncologists, found broad clinical enthusiasm for AI assistance in information synthesis, guideline retrieval, and documentation, but strong consensus that AI should remain advisory rather than autonomous. The fear of #clinician_deskilling — where repeated reliance on AI gradually erodes the independent diagnostic capabilities of oncologists — was raised as a significant concern by experienced clinicians. 4.4 Patient-Facing Applications and Health Literacy The impact of #AI_tools in oncology extends beyond the clinician's consultation room. Chen et al. (2025), in a narrative review published in BMJ Oncology, described the emerging category of patient-facing LLMs capable of disseminating accessible cancer information, answering questions about treatment options in plain language, and providing psychosocial support. This represents a qualitative democratisation of health information that aligns with broader commitments to #patient_empowerment and informed consent. However, the democratisation narrative must be held against its structural constraints. From a world-systems perspective, the ability to benefit from patient-facing AI is conditioned on digital literacy, device access, language coverage, and healthcare infrastructure. In peripheral nations and among marginalised populations in core nations, these conditions are unevenly distributed. #AI_generated health information that is accurate for patients in clinical trial settings may be misleading for those in settings where those treatments are unavailable or unaffordable. 5. Findings 5.1 LLMs Augment but Do Not Replace Clinical Expertise The strongest and most consistent finding across the studies reviewed is that LLMs function best as augmentative tools rather than autonomous decision-makers. No study found that any current LLM, deployed without structured human oversight, produced recommendations of sufficient quality and consistency to replace the expert judgement of a trained oncologist. Benary et al. (2023) established this finding clearly, and subsequent research has refined rather than reversed it. At the same time, the augmentative value is genuine. When LLMs are integrated into tumour board workflows, they accelerate the synthesis of relevant literature, surface treatment options that human annotators may overlook, and provide a first-pass organisation of complex multimodal data that reduces cognitive load for clinicians. Derbal (2025), framing cancer treatment as a problem of controlling a nonlinear dynamical system, argued that generative AI's capacity for data-driven prediction of disease state and treatment response dynamics positions it as a catalyst — not a replacement — for effective oncological decision-making. 5.2 RAG Architecture Is Superior to Standalone LLMs A clear and consistent finding is that standalone LLMs — queried without access to curated, current knowledge sources — perform substantially worse than RAG-augmented systems that ground responses in trusted oncology databases. This finding has direct implications for clinical deployment: institutions considering LLM integration should not deploy general-purpose chat systems directly into clinical workflows but should instead invest in RAG pipelines that connect language models to validated, regularly updated knowledge bases. The Woollie model, described by Heydari et al. (2025) in npj Digital Medicine, exemplifies this approach. Fine-tuned using radiology impression notes from a major cancer centre and externally validated on oncology datasets, Woollie demonstrates that domain-specific fine-tuning combined with curated training data can achieve both high accuracy and genuine scalability. The model prioritises data accuracy, guards against catastrophic forgetting, and establishes a foundation for equitable LLM deployment in oncology. 5.3 Hallucination and Bias Remain Unresolved Challenges #AI_hallucination — the generation of plausible-sounding but factually incorrect clinical information — is the single most serious technical limitation of current LLMs in clinical settings. In oncology, where an incorrect treatment recommendation can result in patient harm, hallucination is not merely a nuisance but a patient safety concern. Multiple studies reviewed here identified hallucination as a primary barrier to clinical adoption, and none reported a reliable solution. Bias is a related but distinct concern. LLMs trained predominantly on English-language, Western clinical literature will encode the demographic and epidemiological biases present in that literature. When deployed in settings with different patient populations, they may produce recommendations that are subtly but significantly less appropriate. Ghebrehiwet et al. (2024), reviewing generative AI in personalised medicine, identified the limitation of foundation models in digital diagnostics, particularly for patient populations underrepresented in training data. From a Bourdieuisian perspective, this is a form of symbolic violence: the imposition of medical knowledge produced in high-capital contexts as though it were universal, without acknowledgement of its situated origins. The patients most likely to be harmed by biased AI recommendations are those already most disadvantaged within the healthcare field. 5.4 Institutional Adoption Is Outpacing Evidence Driven by the three mechanisms of institutional isomorphism — coercive regulatory signals, mimetic competition between elite cancer centres, and normative endorsement from professional oncology societies — the institutional adoption of #AI_tools is proceeding at a pace that exceeds the available clinical evidence. As Wang et al. (2025) noted, most AI-MTB concordance studies are retrospective and conducted at single centres, limiting generalizability. Rigorous prospective, multicentre evaluation of LLMs in real oncology workflows remains rare. Kim et al. (2023), in a multi-organisational qualitative study of AI adoption in US healthcare systems, identified eight key decision points in the institutional AI adoption lifecycle, many of which relate to governance, transparency, and accountability. The study found that health system leaders frequently adopted AI technologies without adequate governance structures in place, replicating the pattern Currie (2012) identified in EHR implementation. For world-systems theory, this trajectory is particularly concerning for peripheral healthcare systems. When AI adoption in oncology is driven by institutional prestige rather than evidence, the technologies that spread globally are those developed and validated for core-nation patient populations, leaving periphery institutions to adopt tools whose safety and effectiveness for their patients has not been established. 5.5 Genomic Data Interpretation and Precision Matching One of the most promising specific applications of LLMs in oncology is the interpretation of #genomic_data for treatment matching. The complexity of next-generation sequencing reports — which may identify dozens of mutations, many of uncertain clinical significance — makes this an ideal domain for AI assistance: the volume of relevant literature exceeds what any clinician can stay current with, and the consequences of missing a clinically actionable variant are severe. Ding et al. (2024), presenting OncoGPT at the Journal of Clinical Oncology, demonstrated that an AI system integrating NGS data analysis with mutation-treatment matching algorithms achieved a weighted F1 score of 0.926 for tissue-of-origin classification across eight cancer types, and significantly improved survival outcomes (hazard ratio 0.326) for patients matched using the system. Chen et al. (2024), presenting the ACMA system, reported that LLM-based treatment matching using real-world data from over 100,000 clinical cases improved median survival rates and achieved a 90% relevance match for clinical trial enrolment. These results, while preliminary and drawn from conference abstracts rather than peer-reviewed studies, point toward a future in which #AI_systems become standard components of the precision oncology workflow. 6. Conclusion The integration of #generative_AI and #large_language_models into oncology is no longer a speculative future — it is a present-tense transformation whose contours are only beginning to come into focus. The evidence reviewed in this article supports a measured but genuinely optimistic assessment. LLMs cannot yet replace the clinical judgement of experienced oncologists, but they can meaningfully extend it: surfacing treatment options that might otherwise be missed, synthesising biomedical literature at scales beyond human capacity, and providing real-time decision support during tumour board deliberations. The Benary et al. (2023) study remains foundational in establishing both the promise and the ceiling of current LLM capability. Its finding that LLMs identified useful treatment options inaccessible to the human expert — while also producing recommendations of substantially lower overall quality — captures the essential paradox of #AI_augmentation in clinical medicine. It is precisely because LLMs can see different things, not better things, that their value as collaborative tools is real. The theoretical frameworks employed in this article illuminate dimensions of #AI_integration that purely technical analyses tend to obscure. Bourdieu's field theory reveals how the adoption of AI is structured by professional capital, habitus, and institutional hierarchy — and why resistance to these tools is not simply ignorance but a rational response to perceived threats to professional authority. World-systems theory forces attention to the global inequities that #AI_development in wealthy nations may reproduce or deepen. Institutional isomorphism explains why adoption is outpacing evidence, driven by legitimacy-seeking behaviour rather than demonstrated clinical benefit. For responsible integration, this article recommends five principles. First, LLMs in oncology should be deployed within RAG architectures that ground outputs in curated, validated knowledge bases, not as standalone chat systems. Second, all clinical LLM deployments should include human-in-the-loop oversight, with explicit governance frameworks that specify when and how AI recommendations can influence clinical decisions. Third, evaluation standards must be established and applied prospectively, across multiple centres and patient populations, before any LLM tool is considered ready for routine clinical use. Fourth, the training data and evaluation populations of oncological AI tools must be diversified to reduce the systematic biases that world-systems analysis identifies. Fifth, the healthcare community must resist the mimetic and coercive pressures of #institutional_isomorphism that encourage premature adoption, and insist that clinical safety evidence precede widespread deployment. The future of #AI_in_oncology is not the replacement of the oncologist, but the creation of a new kind of clinical partnership — one in which human expertise and machine capability are combined in ways that are greater than either alone. Getting that partnership right requires not only better models, but better institutions, better governance, and a more equitable global distribution of the technologies that will define cancer care in the coming decades. This is a call for science to be guided not only by what is technically possible, but by what is socially just. References Benary, M., Wang, X., Schmidt, M., Soll, D., Hilfenhaus, G., Nassir, M., Sigler, C., Knödler, M., Keller, U., Beule, D., Keilholz, U., Leser, U., & Rieke, D. (2023). Leveraging large language models for decision support in personalized oncology. JAMA Network Open, 6(11), e2343689. https://doi.org/10.1001/jamanetworkopen.2023.43689 Benson, R., Elia, M., Hyams, B., Chang, J., & Hong, J. C. (2024). A narrative review on the application of large language models to support cancer care and research. Yearbook of Medical Informatics, 33, 1–12. https://doi.org/10.1055/s-0044-1800726 Benson, R., Kenny, C., Ganjouei, A. A., Zhao, M., Darawsheh, R., Qian, A., & Hong, J. C. (2025). Large language models in population oncology: a contemporary review on the use of LLMs to support data collection, aggregation, and analysis in cancer care and research. JCO Clinical Cancer Informatics, 9, e2500112. https://doi.org/10.1200/CCI-25-00112 Büyükceran, E. U., Seyfettin, A., Babatürk, A., Özkan, M. B., Çolak, D., Ünal, İ., Kaymaz, E., Ergün, E., Emер, M. Ö., & Mersin, H. (2025). High concordance between GPT-4o and multidisciplinary tumor board decisions in breast cancer: a retrospective decision support analysis. Journal of Medical Systems, 49(1), 1–10. https://doi.org/10.1007/s10916-025-02314-9 Chen, D., Parsa, R., Swanson, K., Nunez, J. J., Critch, A., Bitterman, D., Liu, F. F., & Raman, S. (2025). Large language models in oncology: a review. BMJ Oncology, 4(1), e000759. https://doi.org/10.1136/bmjonc-2025-000759 Chen, H., Xu, Z., Zhang, P., Pang, F., Chen, L., Yuan, S., & Wang, K. (2024). Precision oncology and AI: the ACMA system for personalized treatment matching using real-world data. Journal of Clinical Oncology, 42(16 Suppl.), e23134. https://doi.org/10.1200/jco.2024.42.16_suppl.e23134 Currie, W. (2012). Institutional isomorphism and change: the national programme for IT – 10 years on. Journal of Information Technology, 27(3), 236–248. https://doi.org/10.1057/jit.2012.18 Das, U. (2025). Transforming precision medicine through generative AI: advanced architectures and tailored therapeutic design for patient-specific drug discovery. ChemistrySelect, 10(18), e202502448. https://doi.org/10.1002/slct.202502448 Derbal, Y. (2025). Generative artificial intelligence as a catalyst for effective cancer treatments. AI in Clinical Medicine, 3, 1–18. https://doi.org/10.14740/aicm5 Ding, S. D., Xing, X. M., Ding, Y., & Huang, Q. (2024). OncoGPT: an AI assistant for genomic-driven precision oncology. Journal of Clinical Oncology, 42(23 Suppl.), 160. https://doi.org/10.1200/jco.2024.42.23_suppl.160 Erdat, E. C., Yalçıner, M., Örüncü, M. B., Ürün, Y., & Senler, F. (2025). Assessing the accuracy of the GPT-4 model in multidisciplinary tumor board decision prediction. Clinical and Translational Oncology, 27(5), 1–10. https://doi.org/10.1007/s12094-025-03905-1 Ferber, D., El Nahhas, O. E., Wölflein, G., Wiest, I., Clusmann, J., Leßmann, M., Foersch, S., Lammert, J., Tschochohei, M., Jaeger, D., Salto-Tellez, M., Schultz, N., Truhn, D., & Kather, J. (2025). Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nature Cancer, 6(4), 621–634. https://doi.org/10.1038/s43018-025-00991-6 Ghebrehiwet, I., Zaki, N., Damseh, R., & Mohamad, M. S. (2024). Revolutionizing personalized medicine with generative AI: a systematic review. Artificial Intelligence Review, 57(6), 1–38. https://doi.org/10.1007/s10462-024-10768-5 Hamamoto, R., Koyama, T., Takahashi, S., Yasuda, T., Kobayashi, K., Akagi, Y., Kouno, N., Sudo, K., Hirata, M., Sunami, K., & Yamamoto, N. (2026). Implementing generative artificial intelligence in precision oncology: safety, governance, and significance. Journal of Hematology and Oncology, 19(1), 1–25. https://doi.org/10.1186/s13045-026-01781-y Heydari, K., Enichen, E. J., Li, P. C., & Kvedar, J. C. (2025). Incorporating large language models as clinical decision support in oncology: the Woollie model. npj Digital Medicine, 8(1), 1–5. https://doi.org/10.1038/s41746-025-01941-3 Jun, H., Tanaka, Y., Johri, S., Camp, S., Bao, E. L., Carvalho, F. L., Gui, D. Y., Jordan, A. C., Labaki, C., Martin, S. A., Nagy, M., O'Meara, T. A., Pappa, T., Pimenta, E., Saad, E., Yang, D. D., Gillani, R., Tewari, A., Reardon, B., & Van Allen, E. V. (2026). A context-augmented large language model for accurate precision oncology medicine recommendations. Cancer Cell, 44(2), 1–15. https://doi.org/10.1016/j.ccell.2025.12.017 Kehl, K. L. (2025). Use of large language models in clinical cancer research. JCO Clinical Cancer Informatics, 9, e2500027. https://doi.org/10.1200/CCI-25-00027 Kim, J., Boag, W., Gulamali, F., Hasan, A., Hogg, H., Lifson, M. A., Mulligan, D., Patel, M. R., Raji, I. D., Sehgal, A., Shaw, K., Tobey, D., Valladares, A., Vidal, D., Balu, S., & Sendak, M. (2023). Organizational governance of emerging technologies: AI adoption in healthcare. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1–18). ACM. https://doi.org/10.1145/3593013.3594089 Kreimeyer, K., Canzoniero, J., Fatteh, M., Anagnostou, V., & Botsis, T. (2024). Using retrieval-augmented generation to capture molecularly-driven treatment relationships for precision oncology. Studies in Health Technology and Informatics, 316, 1–7. https://doi.org/10.3233/SHTI240575 Nardone, V., Marmorino, F., Germani, M., Cichowska-Cwalińska, N., Menditti, V., Gallo, P., Studiale, V., Taravella, A., Landi, M., Reginelli, A., Cappabianca, S., Girnyi, S., Cwalinski, T., Boccardi, V., Goyal, A., Skokowski, J., Oviedo, R. J., Abou-Mrad, A., & Marano, L. (2024). The role of artificial intelligence on tumor boards: perspectives from surgeons, medical oncologists and radiation oncologists. Current Oncology, 31(9), 5021–5038. https://doi.org/10.3390/curroncol31090369 Pinninti, R., Gullapalli, R., Mallavarapu, K. M., Singareddy, R., Nekkanti, S. S., Nikhil, P., Shukla, S., Hariharan, N., Gudipudi, D., Suseela, K., Koppula, V., Rao, T., & Rajappa, S. (2026). Comparing artificial intelligence and multidisciplinary tumor board decision making in real-world cancer care: a prospective blinded concordance study. ESMO Real World Data and Digital Oncology, 3, e100688. https://doi.org/10.1016/j.esmorw.2026.100688 Rieke, D. T., Wang, X., Schmidt, M., Soll, D., Hilfenhaus, G., Nassir, M., Sigler, C., Knoedler, M., Keilholz, U., Leser, U., & Benary, M. (2023). AI-supported treatment decisions for precision oncology. Annals of Oncology, 34(Suppl. 4), S1536–S1536. https://doi.org/10.1016/j.annonc.2023.09.2869 Shah, Z., Afridi, S. S., Ombada, M., Roy, A. M., LeVee, A., Premji, S., Gupta, V., Lopetegui, N. M., Quiroga, D. M., Sacks, R. L., Shaikh, S., Abdou, Y., Yu, H., Madabhushi, A., Parikh, R., Chaudhary, L. N., Levine, E., Lambertini, M., Kalinsky, K., Kabraji, S., & Gandhi, S. (2025). Benchmarking large language models for clinical decision support in breast cancer care: a multi-institutional expert evaluation. Clinical Cancer Research, 32(4 Suppl.), Abstract PS3-04-06. https://doi.org/10.1158/1557-3265.sabcs25-ps3-04-06 Wang, X., Wang, Q., Ding, G., Wang, J., Tang, Y., & Feng, Y. (2025). Artificial intelligence in multidisciplinary tumor boards enhancing decision making and clinical outcomes in oncology. iScience, 28(6), 114082. https://doi.org/10.1016/j.isci.2025.114082 Yum, J., Naqvi, S. A. A., Zhou, B., & Riaz, I. (2025). Reimagining cancer care with generative artificial intelligence: the promise of large language models. JCO Clinical Cancer Informatics, 9, e2500134. https://doi.org/10.1200/CCI-25-00134 Hashtags #generative_AI #large_language_models #oncology #clinical_decision_support #precision_medicine #personalized_cancer_care #tumor_board_AI #AI_in_healthcare #molecular_oncology #LLM_oncology #retrieval_augmented_generation #AI_ethics_healthcare #institutional_isomorphism #AI_hallucination #cancer_genomics #AI_decision_making #multidisciplinary_oncology #digital_health_equity #Bourdieu_medicine #precision_treatment_matching

University engineering programs have long been criticized for teaching theory in a way that is cut off from the actual work that engineers do. The Conceive–Design–Implement–Operate (#CDIO) framework, set out in detail by Crawley and colleagues (2014), was built to close this gap by placing engineering fundamentals inside the full lifecycle of real products, processes, and systems. This article evaluates the framework and asks how far it has succeeded in modernizing engineering education to meet industrial and innovation demands. Using an integrative literature review, the study draws together recent work on curriculum reform, graduate skills, and the global spread of the framework across more than one hundred member institutions. The analysis is read through three sociological lenses: Bourdieu's concepts of cultural capital and habitus, #world_systems_theory, and institutional isomorphism. The findings suggest that the framework reconnects curricula with practice through active learning, integrated design–build projects, and outcome-based assessment, and that it strengthens the professional skills that employers say are missing. At the same time, the global uptake is shaped less by proof of results than by pressure to look legitimate, which can lead to surface-level adoption. The framework also raises questions about #equity and about academic dependency between core and peripheral nations. The article concludes that the framework is a strong scaffold rather than a neutral technical fix, and that its real impact depends on #faculty_development, resources, and thoughtful local adaptation. Keywords: CDIO framework; engineering education; curriculum reform; institutional isomorphism; cultural capital; world-systems theory; graduate employability; Industry 4.0 1. Introduction For most of the twentieth century, university engineering education followed a familiar pattern. Students spent their early years on mathematics and basic science, moved on to discipline-specific theory, and only met something resembling real practice in a final-year project or once they started a job. This model produced graduates who were strong on analysis but often weak on the practical, social, and #professional_skills that working engineers use every day. Employers across many countries have repeated the same complaint for decades: new graduates know the equations but struggle to design, to build, to work in teams, and to communicate (Souppez, 2025; Rosas Quintero & Namuche Maldonado, 2024). The #CDIO framework was created as a direct response to this problem. It began around the year 2000 as a partnership between the Massachusetts Institute of Technology and three Swedish universities, and it has since grown into a worldwide network. The central idea is simple to state. Engineers in practice conceive ideas, design solutions, implement them, and then operate the resulting systems across their whole life. If that is what engineers do, the argument goes, then engineering education should be organized around the same conceive–design–implement–operate cycle. Crawley and colleagues (2014) present the mature version of this thinking, including twelve standards and a detailed syllabus that lists the knowledge, skills, and attitudes a modern engineer needs. What makes the framework worth studying is not only its pedagogy but its reach. It has been adopted by a large and growing set of institutions on every continent, and it has been written into national reform efforts, accreditation conversations, and quality assurance systems. This raises a question that goes beyond the classroom. When hundreds of universities in very different countries all move toward the same model of curriculum reform, what is actually driving that convergence? Is it evidence that the model works, or is it the pull of legitimacy and the fear of being left behind? This article evaluates the framework on two levels at once. On the first level, it asks a practical question: does it genuinely modernize curricula and produce graduates who are better matched to real-world industrial and #innovation demands? On the second level, it asks a critical question: how should we understand the rapid, near-uniform global spread of the framework, and who benefits from it? To answer the second question, the study uses three well-established sociological tools. #Institutional_isomorphism helps explain why so many universities adopt such similar reforms. Bourdieu's ideas about cultural capital and habitus help explain what kind of engineer the framework tries to create and who is advantaged by that. #World_systems_theory helps explain the flow of educational models from a small set of powerful nations to the rest of the world. The aim is not to praise or to dismiss the framework, but to give a balanced and honest reading of its strengths, its limits, and its politics. The article is organized as follows. Section 2 sets out the background and the three theoretical lenses. Section 3 describes the review method. Section 4 analyzes how the framework modernizes curricula and reads that evidence through the theory. Section 5 reports the main findings. Section 6 concludes with implications for policy and practice. 2. Background and Theoretical Framework 2.1 The CDIO framework in brief The framework rests on a clear claim: that the right setting for engineering education is the practice of engineering itself (Crawley et al., 2014). Rather than treating design and teamwork as extras bolted on at the end, it asks programs to thread them through the whole degree. The framework is expressed through twelve standards that cover the program philosophy, the structure of the curriculum, design–build experiences, the learning workspaces, teaching and #assessment methods, and the development of staff. A companion document, the #CDIO_Syllabus, breaks down the target outcomes into technical knowledge, personal and professional skills, interpersonal skills such as teamwork and communication, and the ability to build whole systems in an enterprise and societal context. A few features set the framework apart from older reform efforts. First, it is built around #integrated_learning, where technical content and skills are taught together instead of in separate silos. Second, it is deliberately #non_prescriptive: it tells a program what to aim for, not exactly how to get there, which allows local adaptation. Third, it links every part of the design to measurable #learning_outcomes, so that programs can check whether they are actually delivering what they promise. In its most recent revision, the framework was extended to give more weight to #sustainability and to optional standards that programs can take on as they mature (Malmqvist, Edström & Rosén, 2020). Recent studies report that combining it with outcome-based education and #project_based_learning improves student engagement and practical competence (Yuan et al., 2024). 2.2 Institutional isomorphism To understand why the framework has spread so widely and so uniformly, the most useful lens is institutional isomorphism. DiMaggio and Powell (1983) observed that organizations in the same field tend to grow more alike over time, not mainly because the shared form is the most efficient one, but because of three kinds of pressure. #Coercive pressure comes from rules, laws, accreditation bodies, and funders. #Mimetic pressure pushes organizations to copy respected peers when they face uncertainty about what to do. #Normative pressure flows from shared training, professional associations, and conferences that spread a common sense of "good practice." In a recent reflection on their original argument, Powell and DiMaggio (2023) note that these mechanisms remain powerful and that fields still drift toward sameness even when actors are trying to innovate. Higher education is a textbook case of an #organizational_field where these forces operate. Studies show that accreditation can act as an isomorphic force that narrows what counts as a good program and discourages real difference, even in markets that reward distinctiveness (Coutet, 2022). The CDIO network itself functions as a field: it has founding members who serve as models, an international set of conferences, shared standards, and a vocabulary that members learn and reproduce. This is exactly the kind of structure that produces convergence. The lens also carries a warning. When organizations adopt a model mainly for legitimacy, they may engage in #decoupling, where the official label is in place but daily practice barely changes. 2.3 Bourdieu: cultural capital, habitus, and field Bourdieu's sociology offers a second lens, one focused on what kind of person an educational model is trying to form and who is helped or harmed in the process. For Bourdieu, a #field is a structured space of competition with its own rules and rewards. People act within a field according to their #habitus, a set of deep dispositions, tastes, and ways of seeing that they pick up through upbringing and schooling. They also carry different amounts of cultural capital, the knowledge, manners, and credentials that a field treats as valuable. Recent scholarship continues to use these tools to explain why educational outcomes track social class so closely (Stahl & Mu, 2022; Jin et al., 2026). Read through this lens, engineering education is a field that does not just transfer knowledge; it builds a professional habitus. The traditional, theory-heavy curriculum rewarded a particular set of dispositions: comfort with abstraction, individual problem-solving, and exam performance. The framework tries to reshape that habitus toward initiative, #teamwork, communication, design judgment, and entrepreneurial thinking. This is a real and arguably positive shift. But Bourdieu's work also points to a risk. The skills the framework rewards, such as confident speaking, leadership, and ease in open-ended design tasks, are often closer to the cultural capital that wealthier students bring with them. A reform that claims to be purely about competence can therefore quietly favor those who already fit, unless programs are careful. Resources matter too: #design_build studios, maker spaces, and laboratories require #economic_capital that not every institution can afford. 2.4 World-systems theory The third lens, world-systems theory, lifts the analysis from single institutions to the global stage. Wallerstein (2004) described the modern world as a single system divided into a powerful #core, a dependent #periphery, and a #semi_periphery in between, with value tending to flow from the edges toward the center. Scholars of higher education have long argued that #knowledge_production follows a similar core–periphery pattern, in which a few wealthy, mostly English-speaking nations set the agenda and the rest adapt to it (Marginson & Xu, 2023; Ge, 2022). The framework fits this map in an interesting way. It was conceived in the core, at MIT and leading Swedish universities, and it has diffused outward to the semi-periphery and periphery, including large adoptions across Asia, Latin America, Africa, and the Middle East. On one reading, this is generous knowledge transfer that helps less-resourced systems leapfrog toward better practice. On a more critical reading, it is a form of academic dependency, in which peripheral universities import a core-designed template and align their graduates to the needs of a global industrial order that the core largely defines. World-systems theory does not force a single verdict here. It also leaves room for upgrading, where countries such as China, Singapore, and South Korea use such models to move up the system rather than to stay fixed at the bottom (Marginson & Xu, 2023). The lens simply insists that we ask whose definition of "real-world demand" the framework is serving. 3. Method This study uses an #integrative_literature_review, a method suited to questions that combine an applied evaluation with a theoretical re-reading. Rather than testing a single hypothesis, an integrative review gathers and synthesizes a diverse body of work, including empirical studies, conceptual papers, and foundational theory, in order to build a fuller picture of a topic and to surface tensions within it. The review was guided by three questions. First, what does the evidence say about how the framework changes engineering curricula and graduate outcomes? Second, how can institutional isomorphism, Bourdieu's field theory, and world-systems theory explain the spread and effects of the framework? Third, what conditions appear to separate genuine #curriculum_reform from #ceremonial adoption? Sources were identified through searches of major academic databases and the published CDIO literature, using combinations of terms such as "CDIO," "engineering education," "curriculum reform," "graduate employability," "Industry 4.0," and the names of the three theories. The study gave priority to peer-reviewed work published within the last five years, while retaining the foundational statements of each theory and the core text by Crawley and colleagues (2014), because these define the concepts under discussion. Items were included if they addressed the framework directly, if they reported evidence on engineering graduate #skills_gaps and #employability, or if they applied one of the three lenses to higher education. Material was excluded if it was promotional, if it lacked a clear method, or if it could not be reliably attributed. The analysis proceeded in two stages. In the first stage, evidence on the framework's pedagogy and outcomes was organized thematically. In the second stage, that evidence was interpreted through each of the three lenses in turn, looking for points where the lenses agreed, where they conflicted, and where they revealed something the framework's own literature tends to leave out. As with any review of this kind, the approach has limits. It depends on published material, which over-represents successful cases and English-language reporting, and the theoretical reading is interpretive rather than statistical. These limits are taken into account when stating the strength of each finding. 4. Analysis 4.1 How CDIO modernizes the curriculum The clearest contribution of the framework is that it reconnects the curriculum to engineering practice. Traditional programs often presented knowledge in an order that made sense to academics but not to learners, leaving students unsure why a given subject mattered. The framework reorganizes the degree around the conceive–design–implement–operate cycle, so that mathematics and science are learned in the service of building real things. This is the heart of #integrated_learning: technical content and professional skills are developed at the same time, in the same projects, rather than in separate boxes. Several pedagogical mechanisms carry this out. #Active_learning replaces passive lecturing with problem-solving, discussion, and hands-on work. #Design_build experiences, often early in the degree, give students the chance to make and test physical or software systems and to learn from failure. #Project_based_learning ties courses together around shared challenges, which helps students see how separate subjects fit. Crucially, the framework insists that #assessment match these goals, so that students are graded on demonstrated competence and not only on written exams. Recent classroom studies that pair it with outcome-based education report gains in engagement, practical ability, and teamwork (Yuan et al., 2024), and broader reviews of engineering competencies for #Industry_4_0 point to exactly the mix of skills the framework targets (Rosas Quintero & Namuche Maldonado, 2024). The framework also responds to the #skills_gap that employers describe. Studies that compare the views of students, academics, and industry find consistent demand for graduates who can communicate, work in teams, and apply knowledge to messy real problems (Souppez, 2025). Because the CDIO Syllabus names these abilities explicitly and requires programs to teach and assess them, it offers a structured way to address complaints that have persisted for years. In this practical sense, the evaluation is positive: the framework is a credible answer to the relevance problem in engineering education. 4.2 An isomorphism reading of the global spread Yet the practical strengths of the framework do not, on their own, explain why so many institutions have adopted it so quickly and in such similar forms. Here #institutional_isomorphism is illuminating. #Coercive pressure is visible wherever national governments, accreditation agencies, and major employers push universities to prove that their graduates are "work-ready," which makes a recognized, outcome-based model attractive. #Mimetic pressure is strong in engineering education because reform is hard and risky; copying respected early adopters such as MIT and leading Swedish schools lowers that risk and borrows their prestige. #Normative pressure runs through the network itself, with its conferences, shared standards, and trained champions who carry a common idea of good practice from one campus to the next (DiMaggio & Powell, 1983; Powell & DiMaggio, 2023). This reading does two useful things. It explains the remarkable convergence of programs that operate in very different national contexts, and it raises a caution that the framework's promotional literature rarely mentions. When the main driver of adoption is the search for #legitimacy, there is a real danger of #decoupling: a university announces that it is a CDIO institution, prints the language in its documents, and satisfies a quality assurance checklist, while the actual teaching changes little. Research on accreditation as an isomorphic force shows that external standards can produce visible compliance and shallow change at the same time (Coutet, 2022). The lesson is not that convergence is bad, but that the label is not proof of substance. 4.3 A Bourdieusian reading of skills and equity Bourdieu's tools turn attention to what the reform does to people. By design, the framework works to replace a narrow academic #habitus with a richer professional one that values initiative, design judgment, collaboration, and communication. For students whose backgrounds already gave them confidence in open, social, project-style settings, this is a comfortable fit. For students whose strength was disciplined individual study, the shift can be harder, and the qualities now rewarded may look a lot like the cultural capital that more advantaged peers bring from home (Stahl & Mu, 2022; Jin et al., 2026). There is also an institutional version of the same point. Design–build studios, #maker_spaces, modern laboratories, and small-group teaching are expensive. Well-funded universities can deliver the full experience, while under-resourced institutions may adopt the standards on paper but cannot supply the workspaces and staff time that make them real. In this way a reform that promises to widen opportunity can, without deliberate counter-measures, reproduce the very #inequality it claims to address. The Bourdieusian point is not an argument against the framework; it is an argument for pairing it with support, mentoring, and investment so that the new habitus it rewards is one all students can actually acquire. 4.4 A world-systems reading of dependency and upgrading Finally, world-systems theory frames the global story. The framework is a #core product that travels to the periphery and semi-periphery, and the "real-world industrial and innovation demands" it serves are, to a large degree, the demands of a global economy whose terms are set in the core (Wallerstein, 2004; Marginson & Xu, 2023). For a peripheral university, adopting it can be a path toward international recognition, mobility for its graduates, and integration into global supply chains and research networks. That is a genuine benefit. The same move, though, can deepen academic dependency. When local programs reshape themselves around a core-designed template and orient their graduates to the priorities of multinational industry, they may sideline local knowledge, local industries, and local development goals. The flow of value, in the form of talent and aligned skills, can run from periphery to core, echoing the broader pattern of the world-system. The more hopeful reading is upgrading: several Asian systems have used imported models to climb toward the core rather than to remain fixed beneath it (Marginson & Xu, 2023). The decisive factor seems to be whether the framework is adopted whole, as a finished import, or adapted, so that its non-prescriptive design is used to serve local needs as well as global ones. 5. Findings Bringing the analysis together, five findings stand out. Finding 1: The framework is an effective tool for modernizing curricula. It reconnects #engineering_education with practice through integrated learning, active learning, design–build projects, and aligned assessment. The available evidence links these methods to better engagement and stronger practical and interpersonal skills, which directly answers the long-standing relevance complaint of employers (Crawley et al., 2014; Yuan et al., 2024; Souppez, 2025). Finding 2: It targets the right skills for current industrial demand. Reviews of the competencies needed for #Industry_4_0 and for the wider skills gap match the abilities named in the CDIO Syllabus, including teamwork, communication, systems thinking, and design ability (Rosas Quintero & Namuche Maldonado, 2024). On its own terms, the framework is well aimed. Finding 3: Global adoption is driven as much by legitimacy as by evidence. The near-uniform spread of the framework across very different countries is best explained by #coercive, #mimetic, and #normative pressures, not only by proof of results. This makes #decoupling a real risk, where the label is adopted without deep change (DiMaggio & Powell, 1983; Powell & DiMaggio, 2023; Coutet, 2022). Finding 4: The reform carries equity questions it does not resolve by itself. Because it rewards dispositions tied to #cultural_capital and depends on costly workspaces, it can reproduce #inequality between students and between institutions unless it is paired with deliberate support and investment (Stahl & Mu, 2022; Jin et al., 2026). Finding 5: Its impact is shaped by core–periphery dynamics. As a core export, the framework can either deepen #academic_dependency or enable #upgrading, depending on whether peripheral institutions import it wholesale or adapt it to local conditions (Wallerstein, 2004; Marginson & Xu, 2023). Taken together, these findings support a measured conclusion. The framework is a strong and coherent answer to a real problem, but its benefits are conditional. Whether a given program achieves genuine #curriculum_reform or merely a ceremonial one depends on #faculty_development, on resourcing, on attention to #equity, and on the willingness to adapt the model rather than copy it. 6. Conclusion The #CDIO framework set out by Crawley and colleagues (2014) is one of the most influential attempts to fix a problem that has dogged engineering education for generations: the distance between what students learn and what engineers actually do. This article has argued that the framework largely succeeds on its own terms. By organizing the curriculum around the #conceive_design_implement_operate cycle, insisting on integrated and #active_learning, building design–build experience into the degree, and aligning assessment with real outcomes, it produces graduates who are better matched to industrial and innovation demands. It names the professional skills employers want and gives institutions a structured way to teach them. But a fair evaluation cannot stop at the brochure. Reading the framework through #institutional_isomorphism shows that its global convergence owes a great deal to the search for #legitimacy, which makes shallow, decoupled adoption a constant danger. Reading it through Bourdieu shows that the new professional #habitus it rewards is bound up with #cultural_capital and with resources, so the reform can widen or narrow #inequality depending on how it is implemented. Reading it through world-systems theory shows that it is a core model travelling to the periphery, capable of enabling #upgrading but also of deepening #academic_dependency. The practical message for universities, accreditors, and policy makers is therefore not "adopt the framework" or "avoid it," but "adopt it well." Genuine impact calls for sustained #faculty_development so that teachers can change how they work, for investment in the workspaces that make design–build real, for explicit steps to ensure that students from every background can build the rewarded habitus, and for the confidence to adapt the non-prescriptive standards to local industries and needs rather than importing them as a finished package. Future research should track programs over time to see how often the label matches deep practice, should study #equity effects directly, and should compare how core and peripheral systems put the framework to work. Treated as a flexible scaffold rather than a fixed blueprint, it can do what it set out to do; treated as a badge, it will mostly produce the appearance of reform. Hashtags #CDIO_framework #engineering_education #curriculum_reform #conceive_design_implement_operate #institutional_isomorphism #cultural_capital #world_systems_theory #graduate_employability #Industry_4_0 #active_learning #design_build_learning #outcome_based_education #STEM_education #RethinkingEngineeringEducation #ModernizingCurricula References Coutet, K. (2022). International school accreditation: An isomorphic force against creativity in a growing competitive market. Journal of Research in International Education, 21(2), 144–160. https://doi.org/10.1177/14752409221117252 Crawley, E. F., Malmqvist, J., Östlund, S., Brodeur, D. R., & Edström, K. (2014). Rethinking engineering education: The CDIO approach (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-05561-9 DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101 Ge, Y. (2022). Internationalisation of higher education: New players in a changing scene. Educational Research and Evaluation, 27(3–4), 229–238. https://doi.org/10.1080/13803611.2022.2041850 Jin, M., Gootjes, D. C., Zhao, H., & Gu, Y. (2026). Family cultural capital and academic achievement: The mediating roles of habitus and field. Frontiers in Psychology, 16, 1745371. https://doi.org/10.3389/fpsyg.2025.1745371 Malmqvist, J., Edström, K., & Rosén, A. (2020). CDIO Standards 3.0 – Updates to the core CDIO standards. Proceedings of the 16th International CDIO Conference. Chalmers University of Technology, Gothenburg, Sweden. Marginson, S., & Xu, X. (2023). Hegemony and inequality in global science: Problems of the center–periphery model. Comparative Education Review, 67(1), 31–52. https://doi.org/10.1086/722760 Powell, W. W., & DiMaggio, P. J. (2023). The iron cage redux: Looking back and forward. Organization Theory, 4(4). https://doi.org/10.1177/26317877231221550 Rosas Quintero, W., & Namuche Maldonado, J. E. (2024). Competencies of the engineer in Industry 4.0 context: A systematic literature review. Production, 34, e20230051. Souppez, J.-B. R. G. (2025). Engineering employability skills: Students, academics, and industry professionals' perception. International Journal of Mechanical Engineering Education. https://doi.org/10.1177/03064190231214178 Stahl, G., & Mu, G. M. (2022). Pierre Bourdieu: Revisiting reproduction, cultural capital, and symbolic violence in education. In B. A. Geier (Ed.), The Palgrave handbook of educational thinkers. Palgrave Macmillan. https://doi.org/10.1007/978-3-030-81037-5_128-1 Wallerstein, I. (2004). World-systems analysis: An introduction. Duke University Press. Yuan, X., Wan, J., An, D., Lu, J., & Yuan, P. (2024). Multi-method integrated experimental teaching reform of a programming course based on the OBE-CDIO model under the background of engineering education. Scientific Reports, 14, 16623.

Clinical documentation has become one of the heaviest hidden tasks in modern medicine, and it is now widely linked to exhaustion, reduced satisfaction, and the slow loss of meaning that clinicians describe when they talk about their work. This article examines how #ambient_AI, the class of tools that listens to a visit and drafts the note for a clinician to review, is being used to ease that load. Taking the quality-improvement survey by Albrecht and colleagues as the focal case, and reading it alongside other recent studies from large health systems, the article asks a simple question with complicated answers: does automating the note actually reduce #clinician_burnout and raise #job_satisfaction, and if so, for whom? The study draws on an integrative review of evidence published between 2021 and 2026, interpreted through three sociological lenses. #institutional_isomorphism explains why hospitals everywhere are adopting these tools at almost the same moment, often before the outcomes are proven. Pierre Bourdieu’s ideas about #cultural_capital and the clinical #field help explain why the note is never only paperwork but a marker of professional standing. #world_systems_theory situates the whole movement inside a global order in which the benefits cluster in well-resourced centres while the data and the dependency spread outward. The evidence suggests real and consistent relief in perceived #documentation_burden and #after_hours_work, with measurable gains in fulfilment, but the gains are uneven and the structural drivers of overwork remain. The article closes by arguing that #ambient_AI is best understood as a redistribution of documentation labour rather than its disappearance. Keywords: ambient artificial intelligence; clinician burnout; clinical documentation; job satisfaction; electronic health records; institutional theory; Bourdieu; world-systems theory 1. Introduction For most of the past fifteen years, the story of the #electronic_health_records system has been told as a story of progress that came with a price. The record made care safer in many ways, but it also turned the clinician into a part-time clerk. Studies have repeatedly shown that doctors and advanced practice providers spend as much time inside the record as they do facing patients, and that a large share of this work spills into the evening, the hours that clinicians now bitterly call #pajama_time. When the day’s visits are over, the typing is not. This quiet shift in what it means to practise medicine has a measurable cost. A 2024 systematic review and meta-analysis pooling sixty-six thousand health-care professionals found that roughly two in five reported burnout, and that those who spent more time on record tasks outside working hours were far more likely to be among them (Wu et al., 2024). #clinician_burnout is not simply tiredness. It is a recognised occupational syndrome built from emotional exhaustion, a growing distance from patients, and a fading sense that the work is effective (Maslach & Leiter, 2022). The #documentation_burden sits close to the centre of that picture, and it has resisted most of the usual fixes, from template libraries to human scribes that were never affordable at scale. Into this gap has come a new class of technology. #ambient_AI tools use #speech_recognition and #natural_language_processing to listen to the conversation between a clinician and a patient, then draft a structured clinical note for the clinician to check, edit, and sign. The clinician keeps authorship and responsibility; the machine takes on the first draft. Within roughly three years these tools moved from small pilots to enterprise deployments at some of the largest systems in the United States, with hundreds of thousands of recorded encounters behind them (Tierney et al., 2024; Tierney et al., 2025). This article uses the quality-improvement survey by Albrecht et al. (2025) as its anchor and reads it against a wider recent literature to ask whether #ambient_AI genuinely lowers #clinician_burnout and lifts what the brief for this study calls operational #job_satisfaction, the day-to-day sense that the work is manageable and worth doing. Rather than treating the technology as a self-evident good, the article places it inside three social theories that explain why it spread so fast, what kind of value the clinical note actually carries, and who is positioned to benefit. The aim is a clear-eyed account that takes the encouraging numbers seriously without mistaking them for the whole story. 2. Background and Theoretical Framework 2.1 The documentation burden and its link to burnout The link between the record and exhaustion is now one of the better-established findings in health-services research. A systematic review of hospital studies concluded that the amount of time spent in the record and its poor usability were the strongest predictors of EHR-related stress, with the intensity of the working environment adding further weight (Alobayli et al., 2023). A narrative review focused on surgery reached a similar verdict, identifying #documentation_burden, after-hours time demands, and the in-basket message load as the recurring drivers of #physician_wellbeing problems across specialties (Karwoski et al., 2025). Newer record-derived metrics, such as the cumulated time it takes a clinician to close their charts, have even been shown to track with burnout directly, which means the burden can now be measured rather than merely described (Shah et al., 2024). Two features of this burden matter for what follows. First, it is uneven in time: much of it is displaced into evenings and weekends, which is exactly the period that erodes recovery and family life. Second, it is heavy in #cognitive_load, the mental effort of holding a patient’s story in mind while simultaneously formatting it for billing, coding, and medico-legal defence. #ambient_AI is interesting precisely because it promises to attack both at once, returning evening hours and lifting the live mental effort of writing while listening. 2.2 Institutional isomorphism: why everyone adopts at once The first puzzle is not whether the tools work but why they spread so uniformly and so quickly. The concept of #institutional_isomorphism, developed by DiMaggio and Powell (1983), offers a sharp answer. Their argument is that organisations facing the same environment come to resemble one another not mainly because resemblance is efficient, but because of three pressures. #coercive_isomorphism comes from rules and dependencies: payer requirements, regulators, and the dominant record vendors whose platforms shape what is even possible. #mimetic_isomorphism comes from uncertainty: when the outcome of a new technology is unclear, organisations copy the choices of prestigious peers as a safe bet. When a flagship academic centre announces that thousands of its clinicians now use ambient documentation, others read that as proof and follow, often before independent evidence arrives. #normative_isomorphism comes from the professions themselves: informatics societies, shared training, conference circuits, and vendor-led communities of practice that turn a particular tool into the expected standard of a serious modern department. Read through this lens, the near-simultaneous adoption of #ambient_AI across academic and community systems looks less like many separate rational decisions and more like a field moving together. This is not a criticism of the technology. It is a warning that adoption can run ahead of proof, and that the strength of the evidence and the speed of the spread are two different things that should not be confused. 2.3 Bourdieu: the note as capital, the clinic as a field If institutional theory explains the spread, Bourdieu helps explain the stakes. In Bourdieu’s thinking, every professional world is a #field, a structured space in which people compete using different forms of capital (Bourdieu, 1986). The clinical note is rarely just information. It is a carrier of #symbolic_capital. A thorough, defensible, well-coded note signals competence to colleagues, protects the clinician in disputes, and converts directly into revenue through billing. The ability to produce such notes quickly and correctly is a form of #cultural_capital, learned over years and quietly used to mark who belongs. This is why the documentation burden cannot be treated as a neutral chore to be deleted. The long hours of #pajama_time are partly the cost of maintaining one’s standing in the field. Bourdieu’s idea of #habitus, the set of dispositions that feel natural to people in a given world, explains why many clinicians keep over-documenting even when no rule requires it. The behaviour is felt as conscientiousness, not as waste. #ambient_AI enters this field as a redistribution of who does the documentation labour and who controls its output. On one reading, it returns time-capital to the clinician and frees attention for the patient. On a more cautious reading, it shifts a portion of professional craft to a vendor’s model, creating a new dependency and a new skill, namely the editing and supervision of machine drafts, that will itself become a marker of competence. The technology does not abolish the capital embedded in the note; it changes the currency. 2.4 World-systems theory: who sits at the centre The third lens widens the frame. #world_systems_theory, associated with Wallerstein (2004), describes a global order divided into a wealthy core, a dependent periphery, and a semi-periphery in between, with value flowing toward the centre. Applied to #digital_health, the pattern is hard to miss. The large deployments, the validated outcome studies, and the capital to absorb subscription costs sit overwhelmingly in high-income, high-resource systems. A recent national analysis of United States hospitals found that AI adoption is sharply clustered, with hotspots and coldspots, and that hospitals in more deprived areas were less likely to implement advanced models (Reddy et al., 2026). Studies from lower- and middle-income settings show adoption constrained by infrastructure and data systems that the core takes for granted (Tran et al., 2025). #world_systems_theory frames operational #job_satisfaction itself as partly a core-country condition: the freedom to redesign a workflow around an expensive tool presumes resources that much of the world does not have. It also flags a flow that runs the other way. Clinical conversations are data, and the value extracted from that data, in the form of improved models, accrues to vendors concentrated in the core. The risk is a widening of #global_health_inequality, in which the relief offered by ambient AI deepens rather than closes the gap between centre and periphery. Holding these three lenses together, the article treats ambient AI as simultaneously a workflow tool, a contest over professional capital, and a node in a global economy of data. 3. Method This is a conceptual and integrative review rather than a study of new primary data. The approach is appropriate when the goal is to interpret an emerging and fast-moving body of evidence through an explicit theoretical frame, and when the relevant studies use mixed designs that do not pool neatly into a single statistic. The work proceeded in four steps. First, a focal case was selected. The quality-improvement survey by Albrecht et al. (2025), which evaluated an ambient AI documentation platform among ambulatory clinicians using paired pre- and post-implementation surveys, was chosen because its outcomes map directly onto the question of this article: perceived #documentation_burden, #after_hours_work, clinician burnout risk, and work #job_satisfaction. Second, a comparative evidence base was assembled. Sources were identified from the major biomedical and health-informatics literature published between 2021 and 2026, prioritising peer-reviewed studies that reported clinician-level outcomes for ambient documentation tools, together with systematic reviews establishing the burnout baseline and recent work on AI adoption patterns. Industry press, vendor white papers, and non-peer-reviewed material were used only for background orientation and were not treated as evidence. Third, the evidence was read through the three theoretical lenses set out above. Each study was examined not only for its reported effect but for what it revealed about the drivers of adoption (#institutional_isomorphism), the professional stakes of the note (#cultural_capital and the #field), and the distribution of benefits across settings (#core_periphery dynamics). Fourth, the findings were synthesised narratively, with attention to convergence and divergence across study designs. Because most available studies are pre-post or observational, with self-selected volunteers and self-reported outcomes, the synthesis foregrounds the direction and consistency of effects rather than precise pooled magnitudes. Where a randomised design exists, its results are weighted more heavily. This method has clear limits, discussed in the conclusion, but it allows the question to be addressed honestly with the evidence that currently exists. 4. Analysis 4.1 The focal case Albrecht et al. (2025) surveyed ambulatory clinicians before and after they began using an ambient AI platform, with a pre-implementation response rate near 52 percent and a post-implementation rate near 74 percent. The headline results were strongly positive on the dimensions that matter here. Clinicians reported that the ease of the documentation #workflow improved markedly after adoption, with an odds ratio close to seven, and that the experience of completing notes improved substantially as well. Most respondents agreed that the tool reduced their documentation burden, cut the time they spent documenting outside clinical hours, lowered their risk of burnout, and increased their job satisfaction. Almost half said they could see an additional patient in a session if needed. Notably, neither the clinician’s specialty nor the number of days they had used the tool significantly changed the pattern of responses, which suggests the perceived benefit was broad rather than confined to a few enthusiasts. These are perceptions, not hard productivity counters, and the study authors are careful about that. But perception is exactly the right outcome for job satisfaction and burnout, which are by definition subjective states. A clinician who feels that the evening is their own again is, on the measures that burnout research uses, less burned out. 4.2 Convergence across the wider evidence The focal case does not stand alone, and its direction is echoed across other recent work. A large implementation at an integrated system reported reduced documentation burden and improved patient communication, and a follow-up after more than a year and over two and a half million encounters estimated time savings of more than fifteen thousand hours for users, the equivalent of roughly eighteen hundred working days (Tierney et al., 2024; Tierney et al., 2025). A pre-post study using a validated #professional_fulfillment index found gains in fulfilment among volunteers over a short trial (Misurac et al., 2025). Multi-site work has linked #AI_scribe use to lower self-reported administrative load and burnout, and qualitative studies describe clinicians feeling more present with patients (Shah et al., 2025; Duggan et al., 2025). The single most important study for tempering optimism is also the most rigorous in design. A randomised trial reported only modest improvements in validated burnout, #cognitive_load, and exhaustion measures, on the order of a seven percent improvement in burnout scores relative to control, and stressed that the findings need confirmation in larger samples (Lukac et al., 2025). The gap between the enthusiastic survey perceptions and the modest randomised effect is the central tension in the entire field, and it deserves to be named rather than smoothed over. 4.3 Reading the gap through theory Why might clinicians report large relief while a randomised trial shows a modest effect? The three lenses each offer part of the answer. Through institutional isomorphism, the enthusiasm itself is partly a field effect. When a tool arrives carrying the prestige of flagship adopters and the framing of inevitable progress, respondents to a non-anonymous post-survey may report benefit in part because the benefit is socially expected. This does not make the relief false, but it cautions against reading high agreement as pure measurement. Through Bourdieu, the modest randomised effect makes sense because the tool addresses the visible draft of the note but not the deeper structure that makes documentation heavy. The billing logic, the medico-legal defensiveness, and the #habitus of over-documentation all persist. The machine writes faster, but the field still rewards the same behaviours, so a portion of the saved time is quietly reabsorbed into more notes, more patients, or more careful editing. Through world-systems theory, the very studies producing these results come from well-resourced cores, which means the reported benefits describe what is possible at the centre, not what is typical across the system as a whole. The relief is real where it is measured, and least available where the burden is arguably worst. 5. Findings Bringing the analysis together, four findings stand out. First, the evidence consistently points in one direction on perception. Across designs and settings, clinicians who use ambient AI tend to report a lighter documentation burden, less #after_hours_work, and higher job satisfaction (Albrecht et al., 2025; Tierney et al., 2024; Misurac et al., 2025; Shah et al., 2025). The consistency of direction across independent groups is itself meaningful, even where individual studies are limited. Second, the size of the effect on validated burnout is real but moderate, not transformational. The most rigorous evidence to date shows improvement measured in single-digit percentages rather than a wholesale cure (Lukac et al., 2025). This positions ambient AI as a genuine contributor to #burnout_prevention rather than a complete solution, and it suggests that systems expecting the technology to fix burnout on its own will be disappointed. Third, the mechanism is best understood as redistribution, not elimination, of documentation labour. The first draft moves from the clinician to the model, but authorship, review, and accountability remain human, and a new skill of supervising machine output emerges. In Bourdieu’s terms, the cultural capital embedded in the note does not vanish; it migrates and changes form. The clinician’s value increasingly lies in judgement and editing rather than in raw transcription. This is a meaningful upgrade in the quality of the work, which likely explains why job satisfaction rises even when measured burnout falls only modestly: the time saved is partly converted into more rewarding attention to patients. Fourth, the benefits are unequally distributed, and the structures producing overwork are untouched. The relief is concentrated in well-resourced, mostly high-income settings, consistent with a #core_periphery pattern and with evidence that AI adoption clusters away from deprived areas (Reddy et al., 2026; Tran et al., 2025). Meanwhile the deeper drivers of documentation burden, including billing complexity, defensive medicine, and message overload, remain in place. Ambient AI lightens the symptom while leaving much of the cause intact, and without deliberate effort it risks widening #global_health_inequality rather than narrowing it. Taken together, these findings support a measured optimism. The technology earns its place as a tool that improves the texture of clinical work and modestly reduces burnout, and clinicians value it. But the same evidence warns against treating it as a structural fix, and against assuming its benefits will reach the settings that need them most. 6. Conclusion The promise of the AI-augmented clinician is appealing because it speaks to a real and well-documented harm. Years of evidence show that the #electronic_health_records system, for all its benefits, helped turn clinicians into clerks and that the resulting documentation burden is tightly bound up with exhaustion and the loss of meaning at work (Wu et al., 2024; Alobayli et al., 2023). Ambient AI offers a credible way to lift part of that load by handling the first draft of the note, and the recent evidence, anchored by Albrecht et al. (2025) and echoed across multiple systems, shows consistent improvement in how clinicians experience their work, alongside a real but moderate reduction in validated burnout (Lukac et al., 2025). The three theoretical lenses used here keep that optimism honest. Institutional isomorphism reminds us that the speed of adoption is not by itself proof of benefit, and that a field can move together for reasons of prestige and pressure as much as evidence. Bourdieu reminds us that the note is a form of capital, so easing the burden is really a redistribution of professional labour and standing, not its disappearance, and that the structures rewarding over-documentation persist. World-systems theory reminds us that the relief is concentrated in the core and that the flow of clinical data outward carries its own risks of dependency and widening inequality. The practical implication is that ambient AI should be implemented as one part of a broader strategy rather than as a standalone rescue. Systems that pair the tool with real reductions in billing complexity, message load, and defensive documentation, and that attend deliberately to equitable access, are the ones most likely to convert modest measured gains into durable improvements in #physician_wellbeing. Systems that adopt it to keep pace with peers while raising patient quotas may find that the saved time simply refills with new work, leaving burnout where it was. The most accurate way to describe the current moment is this. Ambient AI is a genuine and welcome improvement in the daily craft of medicine, valued by the clinicians who use it, and a meaningful contributor to #burnout_prevention. It is not a cure for burnout, because burnout is produced by structures that a transcription tool cannot reach. Recognising both halves of that statement is the difference between using the technology wisely and expecting too much of it. 6.1 Limitations This article is an interpretive synthesis, not a meta-analysis, and it inherits the limits of its sources. Much of the available evidence relies on self-selected volunteers, self-reported outcomes, and pre-post designs without controls, which tend to overstate benefit. Randomised evidence is still scarce. The focal case is a single-platform quality-improvement survey in an ambulatory setting, so its findings may not transfer to inpatient care, to other tools, or to other countries. The theoretical reading offered here is one defensible interpretation among several. These limits do not undermine the central conclusion, but they do mean it should be held provisionally and revisited as larger and more rigorous studies appear. #The_AI_Augmented_Clinician #Ambient_AI_in_Practice #Ambient_AI #Clinician_Burnout #Automated_Documentation #Operational_Job_Satisfaction #AI_Scribe #Clinical_Documentation #Documentation_Burden #Physician_Wellbeing #Electronic_Health_Records #Generative_AI_in_Healthcare #Digital_Health #Healthcare_Innovation #Burnout_Prevention #Institutional_Isomorphism #Bourdieu_in_Healthcare #World_Systems_Theory #AI_in_Medicine #Health_Information_Technology References Albrecht, M., Shanks, D., Shah, T., Hudson, T., Thompson, J., Filardi, T., Wright, K., Ator, G. A., & Smith, T. R. (2025). Enhancing clinical documentation with ambient artificial intelligence: a quality improvement survey assessing clinician perspectives on work burden, burnout, and job satisfaction. JAMIA Open, 8(1), ooaf013. https://doi.org/10.1093/jamiaopen/ooaf013 Alobayli, F., O’Connor, S., Holloway, A., & Cresswell, K. (2023). Electronic health record stress and burnout among clinicians in hospital settings: A systematic review. Digital Health, 9, 1–24. https://doi.org/10.1177/20552076231220241 Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of Theory and Research for the Sociology of Education (pp. 241–258). Greenwood Press. DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101 Duggan, M. J., Gervase, J., Schoenbaum, A., Hanson, W., Chitti, M., Patel, M. S., & Mishuris, R. G. (2025). Clinician experiences with ambient scribe technology to assist with documentation burden and efficiency. JAMA Network Open, 8(2), e2460637. https://doi.org/10.1001/jamanetworkopen.2024.60637 Karwoski, A., Hricz, N., Schlidt, K., & Rasko, Y. M. (2025). Electronic health record burnout in surgery and strategies to improve burnout: A narrative review. Cureus, 17(12), e98469. https://doi.org/10.7759/cureus.98469 Lukac, P. J., Turner, W., Vangala, S., Chin, A. T., Khalili, J., Shih, Y.-C. T., Cheng, E. M., & Mafi, J. N. (2025). Ambient AI scribes in clinical practice — a randomized trial. NEJM AI, 2(12). https://doi.org/10.1056/AIoa2501000 Maslach, C., & Leiter, M. P. (2022). The Burnout Challenge: Managing People’s Relationships with Their Jobs. Harvard University Press. Misurac, J., Knake, L. A., & Blum, J. M. (2025). The effect of ambient artificial intelligence notes on provider burnout. Applied Clinical Informatics, 16(2), 252–258. https://doi.org/10.1055/a-2461-4576 Reddy, S., et al. (2026). The landscape of AI implementation in US hospitals. Nature Health. https://doi.org/10.1038/s44360-025-00016-7 Shah, M., De Arrigunaga, S., Forman, L. S., West, M., Rowe, S. G., & Mishuris, R. G. (2024). Cumulated time to chart closure: a novel electronic health record-derived metric associated with clinician burnout. JAMIA Open, 7(1), ooae009. https://doi.org/10.1093/jamiaopen/ooae009 Shah, S. J., Devon-Sand, A., Ma, S. P., Jeong, Y., Crowell, T., Smith, M., Liang, A. S., Delahaie, C., Hsia, C., Shanafelt, T., Pfeffer, M. A., Sharp, C., Lin, S., & Garcia, P. (2025). Ambient artificial intelligence scribes: physician burnout and perspectives on usability and documentation burden. Journal of the American Medical Informatics Association, 32(2), 375–380. https://doi.org/10.1093/jamia/ocae295 Tierney, A. A., Gayre, G., Hoberman, B., Mattern, B., Ballesca, M., Kipnis, P., Liu, V., & Lee, K. (2024). Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catalyst Innovations in Care Delivery, 5(3). https://doi.org/10.1056/CAT.23.0404 Tierney, A. A., Gayre, G., Hoberman, B., Mattern, B., Ballesca, M., Wilson Hannay, S. B., Castilla, K., Lau, C. S., Kipnis, P., Liu, V., & Lee, K. (2025). Ambient artificial intelligence scribes: learnings after 1 year and over 2.5 million uses. NEJM Catalyst Innovations in Care Delivery, 6(5). https://doi.org/10.1056/CAT.25.0040 Tran, D. M., Dung, N. T., Duc, C. M., Hon, H. N., Khoi, L. M., Hau, N. P., Huyen, D. T. T., Thu, H. T. L., Duc, T. V., Yen, L. M., Thwaites, C. L., & Paton, C. (2025). Status of digital health technology adoption in 5 Vietnamese hospitals: Cross-sectional assessment. JMIR Medical Informatics, 13, e53483. https://doi.org/10.2196/53483 Wallerstein, I. (2004). World-Systems Analysis: An Introduction. Duke University Press. Wu, Y., Wu, M., Wang, C., Lin, J., Liu, J., & Liu, S. (2024). Evaluating the prevalence of burnout among health care professionals related to electronic health record use: Systematic review and meta-analysis. JMIR Medical Informatics, 12, e54811. https://doi.org/10.2196/54811