AI in clinical research: what the evidence supports today

Where AI is already used

Three uses are established enough to appear in mainstream trials: screening electronic records against eligibility criteria to speed recruitment; automated or assisted reading of images and physiological signals for endpoint adjudication; and risk models that flag participants for closer monitoring. In each case a model replaces a repetitive human task and people check its output.

Where the evidence is thin

Most published clinical prediction models have been evaluated only on the data used to build them. External validation, prospective evaluation and studies of clinical impact remain uncommon, and performance often falls when a model meets a new population, a new site or a new scanner. Large language models add a further layer: they are fluent, which makes their errors harder to notice, and they can produce references that do not exist.

What ethics committees and regulators now ask for

Reporting guidance has caught up with practice. TRIPOD+AI covers prediction models, SPIRIT-AI and CONSORT-AI cover trial protocols and reports in which an AI intervention is evaluated, and DECIDE-AI covers early live clinical evaluation. Expect questions about the provenance of training data, the validation strategy, subgroup performance, human oversight, handling of model failure, and how the model version is frozen for the duration of the study.

Language models in evidence synthesis and documentation

Language models can screen literature, summarise protocols and draft study documents quickly, but they invent references and smooth over uncertainty. A practical rule: use them for structure and first drafts, verify every fact and citation against the source yourself, keep a human accountable for each claim, and record in the protocol exactly what was used and for what.

A short checklist before adding AI to a protocol

1. Name the task and the human decision it supports.
2. Specify data sources, the consent basis and where the data are processed.
3. State the validation plan, including external data where possible.
4. Define what happens when the model is wrong or unavailable.
5. Choose the reporting guideline you will follow from day one.

This note is an expert summary. The sources below carry the full evidence base.