Sample PhD Literature Review Chapter: AI Adoption in Healthcare

This is a sample PhD literature review chapter (Chapter 2 of a doctoral thesis) written by an AHC subject expert in Health Informatics to illustrate distinction / doctoral standard. It offers a critical synthesis of published research and does not report new primary data or original empirical findings. Use it as a model for structure, doctoral-level critical argument and Harvard referencing, not as material to submit as your own.

2.1 Introduction and Scope of the Review

Artificial intelligence (AI) has moved, within little more than a decade, from a speculative adjunct to clinical practice towards a set of technologies that health systems are actively procuring, piloting and, in a growing number of cases, embedding in routine care. Yet the striking feature of the field is the persistent distance between what AI systems can be shown to do in controlled studies and what they are actually permitted to do in the daily work of clinicians and organisations. Demonstrations of diagnostic performance that match or exceed specialist clinicians coexist with adoption rates that remain modest, uneven and, in many settings, reversible. It is this gap between demonstrated capability and realised adoption that motivates the present study, and this chapter examines what the existing literature reveals about the conditions under which AI is, and is not, taken up in healthcare.

The purpose of this review is threefold. First, it seeks to establish the theoretical foundations on which any account of AI adoption in healthcare must rest, drawing on the established traditions of information systems (IS) research and the diffusion of innovations. Second, it critically appraises the empirical and conceptual literature on the drivers of and barriers to AI adoption in clinical settings, giving particular attention to the sociotechnical and organisational factors that a narrowly technological reading tends to overlook. Third, it synthesises these strands in order to identify a defensible gap in current knowledge and to position the conceptual framework that guides the remainder of the thesis.

The scope of the review is deliberately bounded. It concentrates on AI as applied to clinical and clinically adjacent tasks — diagnosis, prognosis, decision support, triage and workflow — rather than on the use of AI in biomedical research, drug discovery or the business administration of health organisations, except where those domains illuminate the clinical case. It treats “adoption” not as a single moment of procurement but as a process encompassing acceptance, implementation, routinisation and sustained use, since the literature increasingly indicates that the difficulties of AI lie less in initial acquisition than in durable integration. The review is concerned primarily with the perspectives of clinicians and healthcare organisations, though the interests of patients, regulators and developers are considered where they shape adoption. Geographically, the emphasis falls on high-income health systems, particularly those of the United Kingdom, the United States and comparable European and Commonwealth settings, because these dominate the published evidence; the limits this imposes on transferability are acknowledged explicitly in the synthesis.

The chapter proceeds as follows. Section 2.2 describes the approach taken to identifying and appraising the literature. Section 2.3 examines the theoretical foundations of technology adoption. Section 2.4 surveys the principal clinical applications of AI. Sections 2.5 and 2.6 analyse, respectively, the drivers of and barriers to adoption, with the latter organised around trust, ethics, regulation and workflow. Section 2.7 considers the pivotal role of clinicians. Section 2.8 synthesises the review and articulates the research gap; Section 2.9 sets out the conceptual positioning of the study; and Section 2.10 summarises the chapter.

2.2 Approach to the Review

The review adopts a hybrid strategy that combines the transparency and rigour associated with systematic searching with the interpretive breadth of a narrative, critical review. A fully systematic review in the Cochrane tradition was judged inappropriate to the research question. Such reviews are best suited to answering tightly specified questions about intervention effectiveness where outcomes are homogeneous and comparable; the present question, concerning the conditions of adoption across heterogeneous technologies, settings and stakeholder groups, calls instead for the kind of configurative synthesis that Greenhalgh et al. (2018) argue is necessary where the object of study is complex and context-dependent. A purely narrative review, however, risks selective citation and the privileging of a preferred argument. The hybrid approach was therefore adopted to secure the auditability of a systematic search while retaining the analytical latitude to interpret and connect diverse literatures.

Literature was identified through structured searching of MEDLINE, Scopus, IEEE Xplore, the ACM Digital Library and the AIS eLibrary, chosen to span the clinical, computational and information-systems literatures that intersect in this field. Search strings combined terms for the technology (“artificial intelligence”, “machine learning”, “deep learning”, “clinical decision support”, “predictive algorithm”) with terms for the process of interest (“adoption”, “acceptance”, “implementation”, “diffusion”, “use”) and for the setting (“healthcare”, “clinical”, “hospital”, “physician”, “nurse”). Reference lists of key papers were hand-searched, and forward citation tracking through Google Scholar was used to capture influential recent work not yet fully indexed — a technique known as snowballing that Wohlin (2014) recommends where a field is fast-moving and terminologically unsettled. The search prioritised peer-reviewed journal articles and major conference papers, supplemented by the seminal theoretical works that define the adoption literature and by authoritative grey literature from bodies such as the World Health Organization where it carries particular governance weight.

Appraisal was guided by relevance to the research question and by the methodological quality of each source, but no study was excluded solely on the grounds of design, since the value of qualitative implementation studies and conceptual contributions to this review lies precisely in what randomised or diagnostic-accuracy designs cannot capture. Rather than reporting each study in turn, the synthesis is organised thematically, so that studies are brought into dialogue around the conceptual issues they illuminate. Two limitations of this approach are acknowledged at the outset. First, the pace of publication in AI means that any review is a snapshot; the argument is therefore constructed around durable conceptual claims rather than transient performance figures, which date quickly. Second, publication and reporting biases are acute in this field, where positive demonstrations of algorithmic performance are far more likely to be published than accounts of failed or abandoned implementations (Roberts et al., 2019); the review compensates by weighting implementation and negative-case studies more heavily than their raw frequency in the literature would suggest.

2.3 Theoretical Foundations of Technology Adoption

Any credible account of AI adoption in healthcare must be built on the substantial body of theory that IS and organisational research have developed to explain why individuals and organisations take up, or resist, new technologies. Four traditions are of particular relevance, and their relative strengths and limitations for the AI case structure the discussion that follows.

The first and most influential is the Technology Acceptance Model (TAM), advanced by Davis (1989). Adapting the Theory of Reasoned Action, Davis proposed that the acceptance of a technology is determined chiefly by two beliefs: perceived usefulness, the degree to which a user believes the technology will enhance their performance, and perceived ease of use, the degree to which they believe using it will be free of effort. TAM’s appeal lies in its parsimony and its predictive record across many domains, and it has been applied extensively to clinical technologies including electronic health records and telemedicine (Holden and Karsh, 2010). Yet its parsimony is also its weakness in the present context. TAM locates the explanation of adoption almost entirely within the individual’s cognition, treating the organisational, professional and political context as, at most, an external variable. For a technology such as AI, whose adoption implicates questions of clinical accountability, professional identity and institutional risk, a model that abstracts the user from these structures explains only a fraction of what matters.

The second tradition, the Unified Theory of Acceptance and Use of Technology (UTAUT), represents an attempt to remedy some of TAM’s narrowness. Venkatesh et al. (2003) synthesised eight prior models into four core determinants — performance expectancy, effort expectancy, social influence and facilitating conditions — moderated by variables such as experience and voluntariness of use. The later extension, UTAUT2 (Venkatesh, Thong and Xu, 2012), added constructs including hedonic motivation and habit. UTAUT’s inclusion of social influence and facilitating conditions makes it more sensitive than TAM to the collective and infrastructural dimensions of adoption, and it has been applied productively to health technologies. Nonetheless, it remains fundamentally a variance model concerned with predicting individual intention, and critics have noted that its very comprehensiveness can produce explanatory breadth at the expense of the deeper mechanisms of implementation over time (Bagozzi, 2007). It tells us which factors correlate with intention but comparatively little about how adoption unfolds as a process.

The third tradition addresses precisely this processual dimension. Rogers’s (2003) Diffusion of Innovations theory conceptualises adoption as a temporal process in which an innovation spreads through a social system via communication channels, mediated by the perceived attributes of the innovation — relative advantage, compatibility, complexity, trialability and observability — and by the characteristics of adopters, from innovators through to laggards. Rogers’s framework is valuable for the AI case because it foregrounds the social and communicative nature of adoption and the importance of compatibility with existing values and practices, a consideration acutely relevant to a profession as norm-governed as medicine. Greenhalgh et al. (2004), in an influential systematic review, adapted diffusion theory specifically to the diffusion of innovations in health service organisations, demonstrating that adoption in this sector is rarely a matter of rational individual choice and is instead shaped by organisational readiness, the work of intermediaries and the fit between innovation and system. Their analysis represents a decisive move away from the individualism of TAM towards a systemic and organisational reading.

The fourth and, for this thesis, the most consequential body of work is the explicitly sociotechnical literature that treats technology and the social world as mutually constitutive rather than separable. Here the outstanding contribution is the Non-adoption, Abandonment, Scale-up, Spread and Sustainability (NASSS) framework developed by Greenhalgh et al. (2017). Departing from variance models that seek to predict intention, NASSS is a complexity-informed heuristic that identifies seven interacting domains — the condition, the technology, the value proposition, the adopter system, the organisation, the wider system and the capacity for adaptation over time — and argues that the greater the number of domains exhibiting complexity, the less likely a technology is to be adopted and sustained. NASSS is particularly apt for AI because it directly theorises abandonment and non-adoption as normal outcomes rather than failures to be explained away, and because it insists that sustainability over time, not initial acceptance, is the true test of adoption. Its limitation, which the thesis must confront, is that as a heuristic it is descriptively rich but does not itself generate testable propositions; it organises inquiry rather than predicting outcomes.

Taken together, these traditions reveal a clear intellectual trajectory: from the individual-cognitive parsimony of TAM, through the broader variance modelling of UTAUT, to the processual and systemic accounts of diffusion and sociotechnical complexity. The argument developed later in this chapter is that AI in healthcare exposes the limits of the individual-level models with unusual clarity, because the barriers that most impede AI are not primarily about individual perceptions of usefulness or ease but about trust, accountability, regulation and the reconfiguration of clinical work — matters that only the sociotechnical and organisational traditions are equipped to address. This does not render TAM and UTAUT worthless; they retain value in explaining the micro-level acceptance of specific tools. It does, however, establish the need for a framework that can hold the individual, organisational and system levels together, a need to which the conceptual positioning in Section 2.9 responds.

2.4 Artificial Intelligence Applications in Healthcare

To ground the discussion of adoption, it is necessary to characterise what AI in healthcare actually comprises, since the term encompasses a heterogeneous set of technologies whose adoption dynamics differ markedly. The contemporary wave of clinical AI is dominated by machine learning, and in particular by deep learning using neural networks, which learns statistical patterns from large datasets rather than following hand-coded rules. This represents a decisive departure from the earlier expert systems of the 1970s and 1980s, exemplified by MYCIN, which encoded clinical knowledge as explicit rules and which, despite technical promise, were never widely adopted — a historical precedent whose relevance to the present is considerable and to which the synthesis returns.

The most mature clinical application is in medical image analysis, where the pattern-recognition strengths of deep learning are well matched to the task. Esteva et al. (2017) reported that a convolutional neural network classified skin lesions at a level comparable to that of board-certified dermatologists, while Gulshan et al. (2016) demonstrated high sensitivity and specificity for the detection of diabetic retinopathy in retinal fundus photographs. In radiology, McKinney et al. (2020) reported an AI system for breast cancer screening that, in the conditions of their study, reduced certain categories of error relative to radiologists. These studies are frequently cited as evidence of AI’s readiness for clinical use. A critical reading, however, must resist that inference. As Topol (2019) cautions in his synthesis of the field, and as Kelly et al. (2019) argue in detail, diagnostic-accuracy studies conducted on curated, retrospective datasets establish only that a model can perform a narrow task under favourable conditions; they do not establish that it will maintain that performance prospectively, across the case mix and image quality of routine practice, or that its integration will improve patient outcomes rather than merely matching a benchmark. The distance between in silico accuracy and clinical utility is precisely where adoption succeeds or fails, and it is systematically underreported.

Beyond imaging, a second application domain is clinical prediction and risk stratification. Systems that mine electronic health record data to predict deterioration, sepsis, readmission or mortality have proliferated, and Rajkomar, Dean and Kohane (2019) survey the potential of machine learning to exploit the density of routinely collected data. The cautionary counter-case here is instructive: widely deployed proprietary sepsis-prediction models have, on external validation, performed considerably less well than their developers’ claims implied (Wong et al., 2021), illustrating both the fragility of models transferred across settings and the opacity that commercial deployment can impose on independent scrutiny. A third domain is natural language processing, applied to the extraction of information from clinical narratives and, increasingly, to generative applications such as documentation support; the rapid emergence of large language models has intensified interest here, though the evidence base on their safe clinical use remains immature [VERIFY currency of specific LLM clinical-deployment claims at time of submission]. A fourth domain lies in workflow and operational applications — triage, scheduling, resource allocation — which, while less glamorous than diagnostic AI, may prove more readily adoptable precisely because they do not directly implicate clinical accountability for individual patients.

Two analytical observations emerge from this survey and carry through the chapter. First, the applications differ not only technically but in their adoption profiles: an operational triage tool and an autonomous diagnostic system pose entirely different questions of trust, liability and workflow, and it is an error, common in the enthusiast literature, to speak of “AI adoption” as though it were a single phenomenon. Second, the evidentiary asymmetry noted in Section 2.2 is nowhere more visible than here: the literature is dense with reports of diagnostic performance and sparse in rigorous accounts of sustained clinical deployment. This asymmetry is not incidental to the adoption problem but is itself part of it, since decision-makers are asked to adopt on the basis of evidence that speaks to capability but is largely silent on integration.

2.5 Drivers of Adoption

The forces pressing towards the adoption of AI in healthcare operate at several levels, and the literature, while often assuming rather than examining these drivers, permits their critical reconstruction. At the level of the clinical value proposition, the most cited driver is the promise of improved diagnostic and prognostic performance, particularly in specialties facing rising demand and workforce shortages. Where radiology and pathology confront growing image volumes against constrained specialist supply, AI is presented as a means of extending capacity, and Topol’s (2019) influential review frames the technology’s central promise as the potential to return time to clinicians by absorbing routine analytical labour. This framing, in which AI augments rather than replaces the clinician, recurs throughout the advocacy literature and functions rhetorically to align adoption with professional interests rather than against them.

A second and increasingly prominent driver is systemic and economic. Health systems under fiscal and demand pressure are drawn to AI by the prospect of efficiency: faster throughput, reduced unwarranted variation, earlier intervention that averts costly downstream care, and the automation of administrative burden that is widely implicated in clinician burnout. This driver is powerfully articulated in policy rather than in independent evidence; national strategies, including those of the NHS in England, have positioned AI as instrumental to sustainability [VERIFY specific policy citation and date]. A critical reading must note that the efficiency case is largely prospective and modelled rather than empirically demonstrated at scale, and that the history of health IT — notably the equivocal productivity record of electronic health records (Wachter, 2015) — counsels caution about the assumption that digital tools straightforwardly yield efficiency. The efficiency driver is real as a motivation but insecure as an established fact.

A third driver is competitive and reputational isomorphism. Drawing on DiMaggio and Powell’s (1983) account of institutional isomorphism, one can read the spread of AI initiatives across leading health organisations as driven partly by the mimetic pressure to be seen as innovative and the normative pressure exerted by professional bodies and academic medical centres, rather than solely by demonstrated clinical benefit. This lens helps explain a pattern the enthusiast literature struggles to account for: high rates of pilot initiation coupled with low rates of durable adoption. If a proportion of adoption activity is driven by legitimacy-seeking rather than by value realisation, pilots that satisfy the reputational motive may be allowed to lapse once that motive is met, producing the widely observed phenomenon of “pilotitis” in which promising projects proliferate but few scale (Greenhalgh et al., 2017). A fourth driver is the supply-side push of a well-capitalised technology industry, whose commercial interest in adoption shapes the evidence base, the framing of the technology and the pace of deployment — a factor that the neutral language of “drivers” tends to obscure and that the ethics literature, discussed below, brings sharply into view.

The critical point that unites this analysis is that the drivers of AI adoption are substantially different in kind from the barriers. The drivers are concentrated at the level of the value proposition and the wider system — clinical promise, economic pressure, institutional legitimacy, commercial supply — and are frequently prospective and rhetorical. The barriers, as the next section shows, are concentrated at the level of the adopter, the organisation and the technology-in-use, and are frequently concrete and experiential. This asymmetry, in which abstract system-level enthusiasm meets specific ground-level resistance, is a recurring structural feature of AI adoption and one that individual-level acceptance models are poorly equipped to represent.

2.6 Barriers to Adoption

If the drivers of adoption are relatively diffuse, the barriers are specific, well documented and, on the reading advanced here, decisive. They are analysed under four interlocking headings — trust, ethics, regulation and workflow — though the analysis will show that these are not independent categories but facets of a single sociotechnical problem.

2.6.1 Trust and Explainability

Trust is the pivotal barrier, and it is qualitatively different from the perceived-usefulness construct that dominates acceptance modelling. Asan, Bayrak and Choudhury (2020) argue that clinician trust in AI is a distinct and multidimensional construct shaped by the perceived reliability, transparency and predictability of the system, and that it cannot be reduced to a judgement of accuracy. A clinician may accept that a model is accurate in aggregate yet decline to trust it for the patient in front of them, because clinical accountability is exercised over individuals, not populations. The central technical obstacle to trust is the opacity of contemporary machine learning: deep neural networks are frequently “black boxes” whose reasoning is not inspectable, which sits uneasily with a medical culture that prizes explicit, contestable justification for decisions. Explainable AI has emerged as the proposed remedy, but the literature is sharply divided on whether it resolves the problem. Ghassemi, Oakden-Rayner and Beam (2021) mount a pointed critique, arguing that current post-hoc explanation methods are frequently unreliable and may generate a false sense of understanding that is more dangerous than acknowledged opacity. London (2019) advances the provocative counter-position that medicine already tolerates black boxes — many efficacious drugs act by mechanisms only partly understood — and that the demand for explainability may be misplaced if predictive reliability can be robustly demonstrated. This unresolved debate matters for adoption because it indicates that the trust barrier cannot be dissolved by a technical fix alone; it is at root a question about the grounds on which clinical decisions may legitimately be made.

Trust is further undermined by the fragility of models across contexts. The phenomenon of distributional shift, whereby a model’s performance degrades when the data it encounters in deployment differ from its training data, means that a system validated in one institution may fail silently in another. Cabitza, Rasoini and Gensini (2017) catalogue such unintended consequences, warning additionally of automation bias — the tendency of humans to defer unduly to machine outputs — and of the deskilling that may follow from sustained reliance. These are not reasons to distrust AI in general but reasons why calibrated, context-specific trust is difficult to establish and easy to lose, which is precisely the condition that makes adoption fragile.

2.6.2 Ethics, Bias and Accountability

The ethical barriers to adoption overlap with trust but extend beyond it to questions of justice and responsibility. The most empirically forceful contribution is Obermeyer et al.’s (2019) demonstration that a widely used commercial algorithm for allocating health-management resources exhibited significant racial bias, because it used healthcare cost as a proxy for health need and thereby encoded the effects of unequal access into its predictions. This study is pivotal because it moves the bias concern from the hypothetical to the documented and shows that algorithmic bias arises not from malice but from the ordinary properties of data drawn from an unequal world. Char, Shah and Magnus (2018) generalise the point, arguing that machine learning in healthcare raises ethical challenges at every stage, from the selection of training data to the framing of the prediction task, and that these challenges are frequently invisible to developers focused on technical performance. Mittelstadt et al. (2016), in a widely cited map of the ethics of algorithms, provide the conceptual architecture, distinguishing concerns of inconclusive, inscrutable and misguided evidence from concerns of unfair outcomes and transformative effects on autonomy.

The accountability question is especially corrosive of adoption because it is unresolved in both ethics and law. When an AI-assisted decision causes harm, the locus of responsibility — the clinician who accepted the recommendation, the organisation that deployed the tool, or the developer who built it — is genuinely unclear, and clinicians are understandably reluctant to adopt tools that may expand their liability without expanding their control. Gerke, Minssen and Cohen (2020) analyse this diffusion of responsibility as one of the principal legal obstacles to AI-driven healthcare, and it connects directly to the trust barrier: a clinician asked to bear accountability for an output they cannot inspect faces a rational disincentive to adopt.

2.6.3 Regulation and Governance

Regulatory uncertainty constitutes a third barrier, distinctive to AI because the technology strains regulatory frameworks designed for static products. A machine learning system that continues to learn after deployment is not the fixed device that medical-device regulation presupposes, and regulators have had to develop new approaches to what the field terms “adaptive” or “continuously learning” systems [VERIFY current regulatory framework nomenclature and dates for FDA/MHRA/EU AI Act at time of submission]. The World Health Organization (2021) has issued guidance on the ethics and governance of AI for health that articulates principles — protecting autonomy, ensuring transparency, fostering accountability, promoting equity — but principles are not the operational standards that organisations require in order to adopt with confidence. Reddy et al. (2020) propose a governance model for the clinical deployment of AI, reflecting a recognition that in the absence of settled external regulation, organisations must construct their own assurance mechanisms. The critical implication is that regulatory immaturity does not merely slow adoption administratively; it transfers the burden of assurance onto adopting organisations that are often ill-equipped to bear it, and thereby interacts with the organisational-capacity domain that NASSS identifies as decisive.

2.6.4 Workflow and Organisational Integration

The final barrier, and the one most consistently underestimated in the technical literature, is the integration of AI into clinical workflow. A model that produces accurate outputs delivers no value if those outputs do not reach the right person, at the right moment, in a form that can be acted upon within the constraints of clinical work. The classic sociotechnical analyses of clinical decision support — Sittig and Singh’s (2010) eight-dimensional sociotechnical model, and the long-recognised problem of alert fatigue — establish that integration failures, not analytical failures, account for much of the disappointment of health IT. Sendak et al. (2020), reporting a rare detailed account of implementing a machine learning sepsis tool in practice, demonstrate that success depended less on the model’s accuracy than on months of work to embed it in existing routines, to define who would act on its outputs and how, and to secure the trust of front-line staff. Their account is valuable precisely because it documents the invisible labour of integration that the diagnostic-accuracy literature omits. Coiera (2019) argues more broadly that the guiding metaphor should be human–AI cooperation rather than automation, since the effective unit of performance is the clinician-and-system together, not the system alone. Workflow integration, on this reading, is not a downstream implementation detail but a determinant of whether the technology’s demonstrated capability is ever realised as clinical value.

The cumulative force of this analysis is that the four barriers are not a checklist of separable obstacles but a single interlocking problem: opacity undermines trust; unresolved accountability compounds distrust and is aggravated by regulatory uncertainty; and all three are experienced concretely at the point where the technology meets clinical workflow. This is why, as the synthesis argues, individual-level acceptance models are inadequate to the phenomenon, and why a sociotechnical, multi-level account is required.

2.7 The Role of Clinicians

The clinician occupies a position of unusual centrality in AI adoption, and the literature approaches this centrality from several directions that are worth drawing together critically. At the most basic level, clinicians are the proximate adopters or non-adopters of clinical AI, and their acceptance is therefore a necessary condition of adoption. But to treat them merely as end-users, as the acceptance-modelling tradition tends to, is to miss the deeper ways in which AI implicates professional identity, autonomy and expertise.

Verghese, Shah and Harrington (2018) frame the issue in terms of the enduring value of clinical judgement and the physician’s role, cautioning against a reductive view in which the clinician becomes a mere executor of algorithmic recommendations. Their argument articulates a professional anxiety that is material to adoption: where AI is perceived to threaten the exercise of judgement that clinicians take to be constitutive of their expertise, resistance is not irrational conservatism but a defence of professional values, and it is precisely the incompatibility with existing professional norms that Rogers’s diffusion theory predicts will impede uptake. Conversely, where AI is framed and designed as augmenting rather than supplanting judgement — as decision support subordinate to clinical authority — adoption is more plausible. Shortliffe and Sepúlveda (2018), writing from long experience of clinical decision support, argue that acceptable systems are those that are transparent, that fit the clinical workflow and that leave the clinician in control, effectively specifying the conditions under which the professional will consent to the technology.

A further consideration concerns competence and the changing nature of clinical work. If AI absorbs certain analytical tasks, the concern that clinicians may deskill in those tasks (Cabitza, Rasoini and Gensini, 2017) bears on both safety and professional willingness to cede ground. The role of the clinician is also differentiated: the literature has attended disproportionately to physicians, and particularly to those in image-based specialties, while the perspectives of nurses, allied health professionals and pharmacists — who may be equally or more affected by AI-mediated workflow change — are comparatively neglected, a gap the synthesis records. Finally, clinicians function not only as adopters but as intermediaries and, in Rogers’s terms, as opinion leaders and champions whose endorsement shapes the adoption decisions of colleagues; the implementation literature (Sendak et al., 2020) repeatedly identifies clinical champions as pivotal to success. The clinician is thus simultaneously the gatekeeper, the potential resister and the potential advocate of AI, and no account of adoption that treats them as a passive recipient of a determined technology can be adequate.

2.8 Synthesis and the Research Gap

Drawing the strands together, four propositions emerge from the critical reading of the literature, and their intersection defines the gap this thesis addresses.

First, there is a robust and growing evidence base for the technical capability of clinical AI, but a thin and fragmented evidence base for its sustained adoption in routine practice. The literature is dense at the point of diagnostic-accuracy demonstration and sparse at the point of durable, real-world integration, and this asymmetry is not merely a gap in coverage but a distortion that shapes how decision-makers understand the technology (Kelly et al., 2019; Roberts et al., 2019). The field knows a great deal about what AI can do in principle and comparatively little about the conditions under which it is taken up and kept in use.

Second, the dominant theoretical apparatus applied to adoption is mismatched to the phenomenon. The individual-cognitive models — TAM and, to a lesser degree, UTAUT — that furnish most empirical adoption studies were designed to predict the acceptance of discretionary information systems by individual users, and they locate explanation in perceptions of usefulness and ease. But the analysis of barriers in Section 2.6 shows that the decisive obstacles to AI — trust under opacity, unresolved accountability, regulatory immaturity and workflow integration — are neither individual nor primarily cognitive; they are sociotechnical, organisational and institutional. The sociotechnical and complexity-informed frameworks, above all NASSS (Greenhalgh et al., 2017), are far better matched to the phenomenon, but they are heuristic rather than propositional and have been applied to AI more as descriptive lenses than as the basis for cumulative theoretical development.

Third, the literature is fractured along disciplinary lines. The computational literature theorises the technology but under-theorises its context; the clinical literature evaluates performance but seldom the conditions of use; the IS literature offers adoption theory but has engaged relatively little with the distinctive features of clinical AI; and the ethics and policy literatures illuminate the barriers of trust and accountability but rarely connect them to the empirical dynamics of organisational adoption. No single tradition holds the whole problem, and syntheses that bridge them are scarce.

Fourth, the empirical literature that does exist is skewed in scope: towards physicians rather than the wider clinical workforce, towards image-based diagnostic applications rather than the operational and predictive applications that may adopt more readily, and towards high-income settings whose findings may not transfer. It is also skewed towards the study of intention and initial acceptance rather than of sustained use and, crucially, of abandonment, even though abandonment is, on the NASSS account, a normal and highly informative outcome.

The gap that these propositions jointly define is therefore not a gap in knowledge of AI’s capability, which is abundant, but a gap in integrated, empirically grounded, multi-level understanding of the sociotechnical conditions under which clinical AI is adopted and sustained in routine practice. Specifically, there is a need for research that (a) treats adoption as a process extending to sustainability and abandonment rather than as a moment of acceptance; (b) holds the individual, organisational and system levels together within a single analytical frame rather than privileging any one; (c) foregrounds the sociotechnical barriers of trust, accountability, regulation and workflow that the dominant models marginalise; and (d) attends to the full range of clinical actors and application types rather than to the physician-and-diagnostic-imaging case alone. It is this gap that the present study is designed to address.

2.9 Conceptual Positioning of the Study

The review points towards a clear conceptual positioning. The study rejects the individual-cognitive framing of adoption as insufficient to the phenomenon and adopts instead a sociotechnical, multi-level perspective in which the adoption of clinical AI is understood as the outcome of interactions between the properties of the technology, the dispositions and identities of clinical adopters, the readiness and capacity of organisations, and the wider institutional environment of regulation, reimbursement and professional norms. Within this perspective, the NASSS framework (Greenhalgh et al., 2017) is adopted as the principal organising lens, on the grounds established in Section 2.3: it theorises non-adoption and abandonment as normal outcomes, it spans the levels the phenomenon requires, and it explicitly incorporates the dimension of sustainability over time that the review identifies as under-studied.

NASSS is not, however, treated as sufficient in itself. Because it is heuristic rather than propositional, it is complemented in this study by selective use of the more specified constructs of the acceptance and diffusion traditions — performance and effort expectancy from UTAUT, and compatibility and relative advantage from Rogers — deployed not as a competing model but as a means of giving analytical grain to the adopter and technology domains that NASSS names but does not decompose. The trust construct, developed from Asan, Bayrak and Choudhury (2020) and the explainability debate, is elevated to a central position rather than subsumed under perceived usefulness, reflecting the review’s finding that trust under conditions of opacity and contested accountability is the pivotal and distinctive barrier for clinical AI. In this way the study positions itself at the confluence of the IS adoption, diffusion and sociotechnical traditions, using the sociotechnical frame to hold the whole while drawing on the other traditions for local precision.

This positioning carries methodological implications that are developed in the following chapter but are noted here to make the logic of the thesis explicit. A framing that treats adoption as a context-dependent, multi-level process, and that seeks to understand the sociotechnical conditions of use rather than to predict individual intention, points towards a research design capable of capturing process, context and the perspectives of multiple actors — that is, towards qualitative or mixed-methods inquiry situated in real organisational settings — rather than towards the cross-sectional survey designs that the variance-modelling tradition has predominantly employed. The conceptual positioning and the methodological strategy are thus continuous, each following from the reading of the literature advanced in this chapter.

2.10 Chapter Summary

This chapter has critically reviewed the literature bearing on the adoption of artificial intelligence in healthcare in order to locate and justify the research problem addressed by the thesis. It began by delimiting the scope of the review to the clinical and clinically adjacent applications of AI and to adoption understood as a process extending to sustained use. It set out a hybrid systematic–narrative approach to identifying and appraising the literature, acknowledging the field’s evidentiary asymmetries and biases.

The substantive review established, first, that the theoretical foundations of technology adoption run from the individual-cognitive parsimony of TAM, through the broader variance modelling of UTAUT, to the processual and sociotechnical accounts of diffusion theory and the NASSS framework, and that AI in healthcare exposes the limits of the individual-level models with unusual clarity. It characterised the heterogeneous applications of clinical AI and argued that their adoption profiles differ fundamentally, while noting the systematic over-representation of diagnostic capability and under-representation of sustained deployment in the evidence. It reconstructed the drivers of adoption as diffuse, prospective and concentrated at the system level, and contrasted them with barriers that are specific, experiential and concentrated at the levels of adopter, organisation and technology-in-use. Those barriers — trust under opacity, ethical and accountability concerns, regulatory immaturity and workflow integration — were shown to constitute not a checklist but a single interlocking sociotechnical problem. The pivotal and multivalent role of the clinician, as gatekeeper, potential resister and potential champion, was analysed as central to adoption.

The synthesis identified the research gap not as a deficit in knowledge of AI’s capability but as a deficit in integrated, empirically grounded, multi-level understanding of the sociotechnical conditions under which clinical AI is adopted and sustained. In response, the study was positioned within a sociotechnical, multi-level perspective, adopting the NASSS framework as its principal organising lens, complemented by selective constructs from the acceptance and diffusion traditions and by an elevated treatment of clinician trust. The following chapter develops the methodological strategy that this conceptual positioning entails.

References

Asan, O., Bayrak, A.E. and Choudhury, A. (2020) ‘Artificial intelligence and human trust in healthcare: focus on clinicians’, Journal of Medical Internet Research, 22(6), e15154.

Bagozzi, R.P. (2007) ‘The legacy of the Technology Acceptance Model and a proposal for a paradigm shift’, Journal of the Association for Information Systems, 8(4), pp. 244–254.

Cabitza, F., Rasoini, R. and Gensini, G.F. (2017) ‘Unintended consequences of machine learning in medicine’, JAMA, 318(6), pp. 517–518.

Char, D.S., Shah, N.H. and Magnus, D. (2018) ‘Implementing machine learning in health care — addressing ethical challenges’, New England Journal of Medicine, 378(11), pp. 981–983.

Coiera, E. (2019) ‘The last mile: where artificial intelligence meets reality’, Journal of Medical Internet Research, 21(11), e16323.

Davis, F.D. (1989) ‘Perceived usefulness, perceived ease of use, and user acceptance of information technology’, MIS Quarterly, 13(3), pp. 319–340.

DiMaggio, P.J. and Powell, W.W. (1983) ‘The iron cage revisited: institutional isomorphism and collective rationality in organizational fields’, American Sociological Review, 48(2), pp. 147–160.

Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M. and Thrun, S. (2017) ‘Dermatologist-level classification of skin cancer with deep neural networks’, Nature, 542(7639), pp. 115–118.

Gerke, S., Minssen, T. and Cohen, G. (2020) ‘Ethical and legal challenges of artificial intelligence-driven healthcare’, in Bohr, A. and Memarzadeh, K. (eds.) Artificial Intelligence in Healthcare. London: Academic Press, pp. 295–336.

Ghassemi, M., Oakden-Rayner, L. and Beam, A.L. (2021) ‘The false hope of current approaches to explainable artificial intelligence in health care’, The Lancet Digital Health, 3(11), pp. e745–e750.

Greenhalgh, T., Robert, G., Macfarlane, F., Bate, P. and Kyriakidou, O. (2004) ‘Diffusion of innovations in service organizations: systematic review and recommendations’, The Milbank Quarterly, 82(4), pp. 581–629.

Greenhalgh, T., Wherton, J., Papoutsi, C., Lynch, J., Hughes, G., A’Court, C., Hinder, S., Fahy, N., Procter, R. and Shaw, S. (2017) ‘Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies’, Journal of Medical Internet Research, 19(11), e367.

Greenhalgh, T., Thorne, S. and Malterud, K. (2018) ‘Time to challenge the spurious hierarchy of systematic over narrative reviews?’, European Journal of Clinical Investigation, 48(6), e12931.

Gulshan, V., Peng, L., Coram, M., Stumpe, M.C., Wu, D., Narayanaswamy, A., Venugopalan, S., Widner, K., Madams, T., Cuadros, J., Kim, R., Raman, R., Nelson, P.C., Mega, J.L. and Webster, D.R. (2016) ‘Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs’, JAMA, 316(22), pp. 2402–2410.

Holden, R.J. and Karsh, B.-T. (2010) ‘The Technology Acceptance Model: its past and its future in health care’, Journal of Biomedical Informatics, 43(1), pp. 159–172.

Kelly, C.J., Karthikesalingam, A., Suleyman, M., Corrado, G. and King, D. (2019) ‘Key challenges for delivering clinical impact with artificial intelligence’, BMC Medicine, 17(1), 195.

London, A.J. (2019) ‘Artificial intelligence and black-box medical decisions: accuracy versus explainability’, Hastings Center Report, 49(1), pp. 15–21.

McKinney, S.M., Sieniek, M., Godbole, V., Godwin, J., Antropova, N., Ashrafian, H., Back, T., Chesus, M., Corrado, G.S., Darzi, A., Etemadi, M., Garcia-Vicente, F., Gilbert, F.J., Halling-Brown, M., Hassabis, D., Jansen, S., Karthikesalingam, A., Kelly, C.J., King, D., Ledsam, J.R., Melnick, D., Mostofi, H., Peng, L., Reicher, J.J., Romera-Paredes, B., Sidebottom, R., Suleyman, M., Tse, D., Young, K.C., De Fauw, J. and Shetty, S. (2020) ‘International evaluation of an AI system for breast cancer screening’, Nature, 577(7788), pp. 89–94.

Mittelstadt, B.D., Allo, P., Taddeo, M., Wachter, S. and Floridi, L. (2016) ‘The ethics of algorithms: mapping the debate’, Big Data & Society, 3(2), pp. 1–21.

Obermeyer, Z., Powers, B., Vogeli, C. and Mullainathan, S. (2019) ‘Dissecting racial bias in an algorithm used to manage the health of populations’, Science, 366(6464), pp. 447–453.

Rajkomar, A., Dean, J. and Kohane, I. (2019) ‘Machine learning in medicine’, New England Journal of Medicine, 380(14), pp. 1347–1358.

Reddy, S., Allan, S., Coghlan, S. and Cooper, P. (2020) ‘A governance model for the application of AI in health care’, Journal of the American Medical Informatics Association, 27(3), pp. 491–497.

Roberts, M., Driggs, D., Thorpe, M. et al. (2019) [VERIFY — cited as illustrative of reporting bias in AI diagnostic studies; confirm exact reference and year before submission].

Rogers, E.M. (2003) Diffusion of Innovations. 5th edn. New York: Free Press.

Sendak, M.P., Ratliff, W., Sarro, D., Alderton, E., Futoma, J., Gao, M., Nichols, M., Revoir, M., Yashar, F., Miller, C., Kester, K., Sandhu, S., Corey, K., Brajer, N., Tan, C., Lin, A., Brown, T., Engelbosch, S., Anstrom, K., Elish, M.C., Heller, K., Donohoe, R., Theiling, J., Poon, E., Balu, S., Bedoya, A. and O’Brien, C. (2020) ‘Real-world integration of a sepsis deep learning technology into routine clinical care: implementation study’, JMIR Medical Informatics, 8(7), e15182.

Shortliffe, E.H. and Sepúlveda, M.J. (2018) ‘Clinical decision support in the era of artificial intelligence’, JAMA, 320(21), pp. 2199–2200.

Sittig, D.F. and Singh, H. (2010) ‘A new sociotechnical model for studying health information technology in complex adaptive healthcare systems’, Quality and Safety in Health Care, 19(Suppl 3), pp. i68–i74.

Topol, E.J. (2019) ‘High-performance medicine: the convergence of human and artificial intelligence’, Nature Medicine, 25(1), pp. 44–56.

Venkatesh, V., Morris, M.G., Davis, G.B. and Davis, F.D. (2003) ‘User acceptance of information technology: toward a unified view’, MIS Quarterly, 27(3), pp. 425–478.

Venkatesh, V., Thong, J.Y.L. and Xu, X. (2012) ‘Consumer acceptance and use of information technology: extending the unified theory of acceptance and use of technology’, MIS Quarterly, 36(1), pp. 157–178.

Verghese, A., Shah, N.H. and Harrington, R.A. (2018) ‘What this computer needs is a physician: humanism and artificial intelligence’, JAMA, 319(1), pp. 19–20.

Wachter, R. (2015) The Digital Doctor: Hope, Hype, and Harm at the Dawn of Medicine’s Computer Age. New York: McGraw-Hill Education.

Wohlin, C. (2014) ‘Guidelines for snowballing in systematic literature studies and a replication in software engineering’, Proceedings of the 18th International Conference on Evaluation and Assessment in Software Engineering (EASE ’14). New York: ACM, pp. 1–10.

Wong, A., Otles, E., Donnelly, J.P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M. and Singh, K. (2021) ‘External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients’, JAMA Internal Medicine, 181(8), pp. 1065–1070.

World Health Organization (2021) Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Geneva: World Health Organization.