How to use this list
Large language models can draft text, answer questions and summarise records, and their possible uses in medicine have been reviewed widely (Thirunavukarasu et al., 2023). A systematic review and meta-analysis comparing generative models with physicians found that performance depends on the task, the specialty and the model rather than showing uniform superiority (Takita et al., 2025). For doctoral work this is good news: the open questions are no longer whether a model can produce plausible text, but whether it helps a clinician make a better decision, and how that can be shown.
The rebuilt titles below come from those problem statements, so the page is about generative AI in clinical decision support throughout. Each table gives a possible research question, a likely design and why the topic matters now. For graph models see graph neural networks titles; for transformer-based decision support in neurology, clinical decision support titles; for drug discovery, generative AI in drug discovery; and for the wider field, the topics hub.

Theme 1: Clinicians, trust and appropriate reliance
| Dissertation title | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Clinician trust, workload and workflow fit when a generative decision support tool is introduced in one specialty: a mixed-methods study | How do clinicians in the unit use, question or override a generative decision support tool, and what happens to their documentation workload and confidence? | Mixed methods: observation and interviews alongside routine workload measures before and after introduction, reported to DECIDE-AI | Early clinical evaluation guidance asks for workflow, workload and safety outcomes, not accuracy alone (Vasey et al., 2022) |
| Do tailored explanations change appropriate reliance on generative clinical advice? A randomised reader study with clinicians of differing experience | When explanations are adapted to a clinician's experience level, do clinicians accept more correct advice and reject more incorrect advice than with a single standard explanation? | Randomised reader study using vignettes with known correct answers, including deliberately incorrect model outputs; reliance measured separately for correct and incorrect advice | Explanation methods are proliferating in healthcare, while their effect on decisions is under-studied (Chaddad et al., 2023) |
Evidence that the gap is real. The DECIDE-AI guideline was written because early-stage clinical evaluation of AI decision support was reported inconsistently, and it asks explicitly about human factors, workflow and safety (Vasey et al., 2022). Chaddad et al. (2023) survey explainable AI techniques used in healthcare and show how much work focuses on producing explanations rather than testing what clinicians do with them. Measuring acceptance of correct and incorrect advice separately is what turns "does the clinician trust it?" into a researchable question.
Theme 2: Grounded summarisation and evaluation
| Dissertation title | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Retrieval-grounded summarisation of patient records for clinical handover: accuracy, omission and clinician-rated usefulness | Do retrieval-grounded summaries contain fewer unsupported statements and fewer clinically important omissions than ungrounded summaries, judged against clinician reference summaries? | Retrospective study on de-identified records with ethical approval; blinded clinician rating; error taxonomy covering unsupported statements and omissions; reported to TRIPOD-LLM | Reporting guidance for studies using large language models now exists and is still rarely followed (Gallifant et al., 2025) |
| Privacy-preserving deployment of generative clinical decision support: what is lost when models run inside hospital boundaries? | How much summarisation or question-answering quality is given up when a smaller model is run within the hospital instead of a larger external service, and is the difference clinically material? | Comparative evaluation of locally deployable and externally hosted models on the same de-identified task set; cost and latency measurement; clinician rating of material differences | Health AI governance guidance stresses data protection and accountability alongside performance (World Health Organization, 2021) |
Evidence that the gap is real. Gallifant et al. (2025) set out the TRIPOD-LLM reporting guideline because studies using large language models frequently omit how prompts, data and evaluation were handled, which makes results hard to compare. World Health Organization (2021) sets ethical and governance expectations for AI in health, including data protection and accountability, which is why local deployment is a legitimate research question rather than a technical detail. A study that reports omissions as carefully as it reports accuracy will be more useful than one more leaderboard score.
Theme 3: Prediction, calibration and reporting
| Dissertation title | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Do generative models add predictive value over established clinical prediction models? A comparative study with calibration and transparent reporting | When a generative model's outputs are added to an established risk model for one outcome, does discrimination or calibration improve on patients from a later period? | Retrospective cohort study with temporal validation; comparison with a regression or boosted baseline; calibration curves; reported to TRIPOD+AI | Prediction model reporting guidance now covers machine learning explicitly (Collins et al., 2024) |
Evidence that the gap is real. Collins et al. (2024) updated prediction model reporting guidance for models that use machine learning, asking authors to report how data were split, how missing data were handled and both discrimination and calibration. Many studies of generative models in clinical settings report accuracy on benchmark questions instead, and Thirunavukarasu et al. (2023) note that evaluation in medicine needs to move towards clinically meaningful endpoints. A title that compares against a strong established model, on later patients, answers a question clinicians actually ask.
How to test a title before you commit
Use the general method in how to choose a PhD topic and how to select a PhD topic, or our topic selection support, and screen each idea with the FINER criteria (feasible, interesting, novel, ethical, relevant) (Hulley et al., 2013). For clinical generative AI titles, check four things early.
- Task: name one clinical task (handover summary, triage question, risk estimate); "clinical decision making" is not a task.
- Approvals: ethics approval, data protection review and a data sharing agreement take months; start before the title is fixed.
- Reference standard: decide what counts as a correct output and who judges it.
- Reporting: choose the guideline you will follow (DECIDE-AI, TRIPOD-LLM or TRIPOD+AI) and design backwards from it (Vasey et al., 2022; Gallifant et al., 2025; Collins et al., 2024).
Worked example (illustrative). A student starts from the live title asking whether multimodal generative AI can improve clinical decision-making. No single study can answer that. With a hospital supervisor, she narrows it to nursing handover summaries in one ward, uses de-identified records under approval, compares grounded and ungrounded summaries, and has two senior nurses rate unsupported statements and clinically important omissions in blinded order. The title names its task, its setting, its comparison and its reference standard.
Common mistakes
- Asking whether generative AI "improves clinical decision-making" without naming a task or setting.
- Reporting benchmark accuracy and calling it clinical validation.
- Counting wrong statements but not clinically important omissions.
- Testing explanations for plausibility rather than for their effect on decisions.
- Assuming ethics approval and data access will be quick.
- Copying titles from another field, so the title and the research question no longer match.
Key takeaways
- Evidence on generative models against clinicians varies by task and specialty, so avoid blanket claims (Takita et al., 2025).
- Early clinical evaluation should cover workflow, workload and safety, not accuracy alone (Vasey et al., 2022).
- Studies using large language models now have reporting guidance to follow (Gallifant et al., 2025).
- Prediction claims need calibration and transparent reporting as well as discrimination (Collins et al., 2024).
- Governance, data protection and accountability belong in the design of health AI research (World Health Organization, 2021).
- Screen every shortlisted title for feasibility, novelty, ethics and relevance (Hulley et al., 2013).
