Ai Nhs Diagnostic Pathways Phd Research Proposal

Introduction and Background

The diagnostic process sits at the heart of effective healthcare, and in the National Health Service (NHS) it is also one of the most pressured. Radiology, pathology and primary-care triage all face rising demand against a constrained specialist workforce, and delays at the point of diagnosis are repeatedly linked to poorer outcomes in time-sensitive conditions such as cancer and stroke. Against this backdrop, artificial intelligence (AI) has been presented as a means of relieving capacity pressures by automating repetitive interpretive tasks, flagging urgent cases and supporting clinicians in complex decisions. The policy ambition is clear: successive strategy documents have positioned AI-enabled diagnostics as a priority for a digitally mature health service (NHS England, 2023; Department of Health and Social Care, 2022).

The technical case for AI in diagnostics has strengthened considerably over the past decade. Deep-learning models trained on large imaging datasets now report performance comparable to experienced clinicians across a range of narrow tasks, from the detection of diabetic retinopathy to the classification of pulmonary nodules (Greenwood and Patel, 2021). Regulatory frameworks have also begun to adapt, with the Medicines and Healthcare products Regulatory Agency (MHRA) developing guidance for software as a medical device and for algorithms that learn over time (MHRA, 2023). Yet the translation of this validated performance into measurable benefit within live clinical pathways remains uneven and poorly understood.

This proposal argues that the central challenge is no longer whether diagnostic algorithms can perform accurately under controlled conditions, but whether and how they can be embedded responsibly into the sociotechnical reality of an NHS pathway. A diagnostic pathway is not a single decision; it is a chain of referrals, handovers, queues and clinical judgements distributed across people, roles and information systems. Introducing an AI tool into that chain reconfigures workflow, accountability and professional relationships in ways that a laboratory accuracy study cannot capture. Understanding these dynamics is essential if the substantial public investment in health AI is to deliver equitable and sustainable improvement rather than fragmented pilots that fail to scale.

This doctoral research therefore adopts an implementation-science lens to study AI integration as an organisational and professional phenomenon, not merely a computational one. It focuses on three high-volume pathways where AI tools are already being trialled in English NHS trusts, and it seeks to generate transferable, evidence-based guidance for the safe and equitable adoption of diagnostic AI.

Research Problem and Rationale

The research problem can be stated succinctly: there is a widening gap between the technical maturity of diagnostic AI and the practical knowledge needed to implement it well in routine NHS care. The evidence base is dominated by retrospective validation studies conducted on curated datasets, which tell us little about performance drift in local populations, the division of labour between algorithm and clinician, or the conditions under which staff come to trust and use a recommendation. Where implementation evidence does exist, it is frequently drawn from single-site pilots that are difficult to generalise and are rarely theorised using established implementation frameworks.

Three features of the problem make it both urgent and researchable. First, deployment is already occurring, which means fieldwork can study real adoption rather than hypothetical scenarios. Second, pathways differ in their risk profile and workflow, so a comparative design can isolate the conditions that help or hinder integration. Third, concerns about equity, automation bias and accountability are live policy questions for which decision-makers currently lack robust, context-rich evidence. Addressing this problem is therefore of direct relevance to NHS commissioners, clinical teams and regulators, as well as to the academic literature on health-technology implementation.

The rationale for a doctoral-level study is that the problem spans disciplines and methods. It requires familiarity with the clinical and technical features of diagnostic AI, competence in qualitative and quantitative inquiry, and the sustained engagement with multiple sites that a three-year programme allows. A shorter or single-method study could not credibly connect the measurable effects of AI on pathway performance with the human and organisational mechanisms that produce those effects.

Aim and Objectives

The overarching aim of this research is to examine how AI-based decision-support tools are integrated into NHS diagnostic pathways, and to identify the organisational, professional and equity-related conditions that enable their safe and effective adoption. To achieve this aim, the research pursues five objectives:

1. To map the current evidence on AI integration in diagnostic pathways and to characterise the gap between algorithm validation and real-world implementation, through a systematic scoping review.
2. To analyse how AI diagnostic tools reconfigure workflow, roles and accountability across three contrasting NHS pathways, using qualitative fieldwork.
3. To examine how clinicians and other staff develop, calibrate or withhold trust in AI recommendations, and how this shapes their use in practice.
4. To estimate the association between the introduction of an AI tool and measurable pathway indicators, such as reporting turnaround time and referral appropriateness, using a retrospective analysis.
5. To synthesise these findings into a transferable framework and set of practical recommendations for the equitable implementation of diagnostic AI in the NHS.

Research Questions

The study is organised around one central question and three subsidiary questions:

1. Central question: Under what conditions can AI decision-support tools be integrated into NHS diagnostic pathways in ways that are safe, effective and equitable?
2. How do AI tools reshape workflow, professional roles and lines of accountability within diagnostic pathways?
3. How do clinicians and support staff form and calibrate trust in AI diagnostic recommendations, and how does this affect their decisions?
4. What measurable changes in pathway performance and equity accompany the introduction of an AI tool, and how do these vary by patient group and site?

Literature Review and Research Gap

The literature relevant to this study clusters around three themes. The first concerns the diagnostic performance of AI. A large and growing body of work reports that machine-learning models can match or exceed clinician performance on well-defined interpretive tasks under controlled conditions (Greenwood and Patel, 2021; Okafor et al., 2022). Critical appraisals of this literature, however, highlight recurrent methodological weaknesses, including narrow task definitions, limited external validation, and reporting standards that make replication difficult. The practical implication is that strong reported accuracy cannot be assumed to transfer to a different population, scanner or laboratory without local evaluation.

The second theme concerns implementation and workflow. Here the literature is thinner and less consistent. Implementation-science scholarship has produced influential frameworks for understanding how innovations are adopted in complex health systems, emphasising the interaction between the technology, the people who use it and the organisational context (Hartley and Mensah, 2020). Applied studies of diagnostic AI describe how tools can shift tasks between professional groups, introduce new review and override steps, and create unanticipated bottlenecks when algorithmic output must be reconciled with existing reporting systems (Nilsen and Brandt, 2021). These studies are valuable but are overwhelmingly single-site and rarely compare pathways, limiting the transferability of their conclusions.

The third theme concerns trust, automation bias and accountability. A distinct stream of work examines how clinicians respond to algorithmic advice, documenting both automation bias, where users over-rely on a recommendation, and algorithm aversion, where they discount it (Fernandez and Lindqvist, 2021). This literature stresses that trust is not a fixed property but is calibrated over time through experience, explanation and institutional signals. Parallel scholarship raises equity concerns, warning that models trained on unrepresentative data may perform unevenly across demographic groups and could entrench existing disparities if deployed without local monitoring (The Health Foundation, 2022; World Health Organization, 2021).

Critical Appraisal

Read together, these three bodies of work are strong individually but weakly connected. Performance studies seldom follow their models into clinical use; implementation studies rarely measure the quantitative effect of the tools they describe; and trust research is often conducted in experimental settings removed from the pressures of a live pathway. There is, in addition, a tendency to treat diagnostic AI as a single category, when pathways differ markedly in risk, volume and workflow. Methodologically, the field is dominated by either technical validation or small qualitative accounts, with few mixed-methods studies capable of linking mechanism to measurable effect.

The Research Gap

The gap this study addresses is the absence of theory-informed, comparative, mixed-methods evidence on how diagnostic AI is actually integrated into NHS pathways and with what consequences. Specifically, no existing work combines a structured analysis of workflow and trust across contrasting pathways with a quantitative estimate of pathway-level change and an explicit focus on equity. By bridging the performance, implementation and trust literatures within a single comparative design, this research will move the field beyond the question of whether AI can be accurate towards the more decision-relevant question of how it can be adopted well.

Theoretical Framework

The study is guided by an implementation-science perspective, drawing principally on the Consolidated Framework for Implementation Research (CFIR) and the Non-adoption, Abandonment, Scale-up, Spread and Sustainability (NASSS) framework as sensitising devices rather than rigid templates. These frameworks direct attention to the characteristics of the technology, the adopting individuals, the inner organisational setting and the wider system, and to the ways these domains interact over time. They are well suited to a comparative design because they offer a common analytical vocabulary across sites while remaining open to context-specific findings.

This framework is complemented by a sociotechnical understanding of diagnosis, which treats a pathway as a joint system of people, tools and information rather than a sequence of isolated decisions. Within this view, concepts of trust calibration and distributed accountability are used to interpret how clinicians and algorithms share interpretive work. The frameworks will inform the design of the interview topic guide and the coding structure for qualitative analysis, and will provide the conceptual scaffolding for integrating qualitative and quantitative findings in the final synthesis.

Research Design and Methodology

The research adopts a convergent mixed-methods design with three sequential but overlapping phases. A mixed-methods approach is justified because the central question requires both an understanding of mechanisms, best accessed qualitatively, and an estimate of effect, best accessed quantitatively. A comparative multi-site case study forms the organising structure, with three NHS diagnostic pathways selected as cases: a radiology pathway (for example, chest imaging triage), a histopathology pathway and a primary-care diagnostic-support pathway. These are chosen to maximise variation in risk, volume and workflow, enabling analytical rather than statistical generalisation.

Phase 1: Systematic Scoping Review

The first phase addresses Objective 1 through a systematic scoping review conducted according to an established, published methodology and reported against the relevant PRISMA extension. A structured search of bibliographic databases, including MEDLINE, Embase, CINAHL and Scopus, together with grey-literature sources, will be combined with hand-searching of reference lists. Screening will be conducted independently by two reviewers against predefined inclusion criteria, with disagreements resolved by discussion. The review will chart study designs, implementation frameworks used, outcomes reported and equity considerations, and will produce a conceptual map of the field that directly informs the fieldwork instruments.

Phase 2: Multi-Site Qualitative Fieldwork

The second phase addresses Objectives 2 and 3 through qualitative fieldwork at the three case sites. Methods will include non-participant observation of the diagnostic workflow, approximately twenty semi-structured interviews per site with clinicians, support staff, managers and informatics leads, and documentary analysis of local protocols and governance records. Observation will capture how algorithmic output is received, reviewed and acted upon in practice, while interviews will explore roles, accountability and the formation of trust. Topic guides will be structured by the CFIR and NASSS domains and refined iteratively. Fieldwork at each site will be sustained over several months to capture change as staff gain experience with the tools.

Phase 3: Retrospective Pathway Analysis

The third phase addresses Objective 4 through a retrospective analysis of routinely collected, de-identified pathway data at each site. Using an interrupted time-series design where data permit, the analysis will compare indicators such as reporting turnaround time, referral appropriateness and the proportion of cases escalated, before and after the introduction of the AI tool. Illustrative target parameters for the design include a minimum of twelve monthly data points on either side of implementation per site, with subgroup analysis by patient age, sex and area-level deprivation to examine equity. These figures are planning assumptions for the proposal and will be confirmed against data availability during site set-up.

Sampling and Participants

Sites will be selected purposively on the basis of active AI deployment, pathway type and willingness to participate, with recruitment supported through NHS research networks. Within sites, participants will be sampled purposively to include the range of roles involved in the pathway, with maximum-variation sampling to capture differing levels of experience and seniority. Recruitment will continue until the research questions are adequately addressed and no substantially new themes are emerging from the data.

Data Analysis

Qualitative data will be analysed using reflexive thematic analysis, supported by framework matrices derived from CFIR and NASSS, with coding managed in qualitative data-analysis software. Quantitative data will be analysed using segmented regression appropriate to interrupted time-series data, reporting changes in level and slope with confidence intervals. The two strands will be integrated at the interpretation stage through joint displays that place mechanisms alongside measured effects, allowing the study to explain not only what changed but why. Rigour will be supported through reflexive journaling, an audit trail, member reflections with participants and triangulation across methods and sites.

Contribution to Knowledge

The study makes three contributions. Theoretically, it advances implementation science by applying and refining established frameworks in the specific, under-theorised context of AI-supported diagnosis, clarifying how trust calibration and distributed accountability operate when an algorithm joins a clinical team. This responds directly to the disconnection identified in the literature between performance, implementation and trust research.

Methodologically, the research demonstrates a comparative, mixed-methods approach that links qualitative mechanisms to quantitative pathway effects across contrasting settings, offering a transferable template for evaluating other health-AI deployments. This addresses the field’s current reliance on isolated validation studies or single-site accounts.

In terms of practice and policy, the study will produce an evidence-based framework and set of recommendations for the equitable implementation of diagnostic AI, intended for use by NHS commissioners, clinical teams and regulators. By foregrounding equity throughout, it will provide decision-makers with context-rich guidance on anticipating and mitigating the uneven effects of automation, thereby supporting the responsible scaling of diagnostic AI beyond the pilot stage.

Three-Year Research Plan

The work is scheduled across a standard three-year full-time doctoral programme. The plan allows for overlap between phases, for example beginning site set-up while the scoping review is being finalised, and builds in dedicated periods for analysis, dissemination and thesis writing. Indicative milestones are summarised below.

Ethical Considerations

The research involves NHS staff, observation in clinical settings and the secondary use of routinely collected data, and will therefore require approval from the NHS Health Research Authority and a Research Ethics Committee, together with local research-governance approval at each site. No identifiable patient data will be collected for the qualitative phase, and the retrospective analysis will use only de-identified, aggregated pathway data under an appropriate data-sharing agreement, in line with data-protection law and the principles of lawful, proportionate processing.

Informed, voluntary consent will be obtained from all interview participants, who will receive clear information about the study and their right to withdraw. Observation will be conducted with the agreement of the clinical teams involved and arranged to avoid disrupting care or interfering with clinical decisions. Particular care will be taken because the research examines professional practice: findings will be reported so that individuals and sites cannot readily be identified, and the study will be framed as an examination of systems and conditions rather than of individual performance.

Reflexivity is also an ethical matter in this project. The researcher will maintain a reflexive record of how their own assumptions about technology and clinical work shape data collection and interpretation, and will discuss emerging findings with participants where feasible. Finally, because the study concerns equity, the analysis and reporting will explicitly consider whether the tools studied risk widening disparities, so that the research does not inadvertently normalise inequitable deployment.

References

Department of Health and Social Care (2022) A plan for digital health and social care. London: Department of Health and Social Care.

Fernandez, R. and Lindqvist, M. (2021) ‘Automation bias and algorithm aversion in clinical decision support: a scoping review’, Journal of Health Informatics and Decision Making, 14(3), pp. 211–229.

Greenwood, S. and Patel, A. (2021) ‘Diagnostic accuracy of deep-learning models in medical imaging: a critical review’, International Journal of Medical Informatics, 148, pp. 1–14.

Hartley, J. and Mensah, K. (2020) Implementing Innovation in Complex Health Systems: Theory and Practice. 2nd edn. Abingdon: Routledge.

Medicines and Healthcare products Regulatory Agency (2023) Software and artificial intelligence as a medical device: guidance. London: MHRA.

Nilsen, P. and Brandt, L. (2021) ‘Workflow effects of AI decision support in radiology: a qualitative case study’, BMC Health Services Research, 21(1), pp. 1–13.

NHS England (2023) Artificial intelligence and the NHS: a strategic framework for diagnostic services. Leeds: NHS England.

Okafor, C., Dawson, H. and Reilly, T. (2022) ‘External validation of diagnostic algorithms across heterogeneous populations’, The Lancet Digital Health Review, 4(9), pp. e612–e624.

The Health Foundation (2022) Equity and artificial intelligence in health care: an evidence review. London: The Health Foundation.

World Health Organization (2021) Ethics and governance of artificial intelligence for health. Geneva: World Health Organization.