AI clinical trial matching: from prescreening to human review

NIH-Developed AI Algorithm Successfully Matches Potential Volunteers to Clinical Trials Release

National Library of Medicine · 2024-11-18 · Source

NLM research announcement published on November 18, 2024, introducing TrialGPT’s approach to trial prescreening. In English; not a current clinical-use authorization or treatment recommendation.

Research: how is a candidate list produced?

The 2024 TrialGPT study combines trial retrieval, eligibility-criterion matching and ranking. Its evaluation includes 183 synthetic patient summaries. In a separate small user study with two physicians, prescreening took 42.6% less time. This does not demonstrate improved real-world enrollment or treatment outcomes; it is a simulated pilot result.

Evidence limits: explanations can also be wrong

The study identifies errors involving reasoning, medical knowledge and label definitions. Adding an explanation to a match does not guarantee that it is correct. Results concern the evaluated tasks and version, and cannot be generalized to every AI tool.

Clinical use: final review belongs to the trial team

NLM describes TrialGPT as an experimental research and discovery tool, not clinical decision support. It does not establish final eligibility or enroll participants. Trial teams must verify current protocols, recruitment status and the relevant site. NLM also calls for only necessary, de-identified summaries. Missing information should not be confused with evidence that a criterion is met.

Regulatory context: draft guidance is not product authorization

FDA’s January 2025 draft guidance discusses context-specific model credibility for AI-generated information supporting regulatory decisions about drugs and biologics. The document is a nonbinding draft; it should not be interpreted as authorization for TrialGPT or any other product.

Editorial proposal: deliver a review record, not just a match

For a team pilot, we suggest separate fields for each candidate trial’s criterion, supporting evidence, missing information and human review. Showing which question remains unanswered is more useful than a single eligibility label. Define success first: review time, missed candidates or incorrect suggestions? These are our editorial proposals, not outcomes measured by the study.

Responsible AI: evaluate access alongside speed

In our editorial assessment, average speed alone is insufficient. Record which errors accumulate across languages and short or incomplete summaries. Provide a feedback route for experts to examine failed matches. This educational article offers no individual trial-participation or treatment advice and makes no claim about the effectiveness of RAI products.