BIGGER PICTURE Better care
A clinical-trial search that explains each eligibility rule
Finding a relevant clinical trial is only the start. Someone still has to compare a person’s history with detailed inclusion and exclusion rules. TrialGPT turns that reading task into an organised, reviewable shortlist.

- Evaluation data
- 183 synthetic patient cases
- Candidate retrieval
- >90% relevant trials in<6% of the collection
- Screening-time result
- 42.6% shorter in a user study
How the idea works
Retrieve possible trials
Patient information narrows a large trial collection.
Check each criterion
The language model compares eligibility rules with the case.
Rank with explanations
A reasoned shortlist supports professional screening.
The difficult part is often the small print
Trials can require a particular diagnosis, prior treatment or laboratory result, while excluding other conditions. Searching for a disease name alone does not settle eligibility. TrialGPT was developed by NIH researchers and collaborators to help organise the patient-to-trial comparison at the level of individual criteria.
Sources: [1]
First narrow the library; then read closely
The system first retrieves plausible trials from a large collection. A second module checks each inclusion or exclusion criterion against the supplied patient information and produces an explanation. A final stage ranks the trials. This separates broad discovery from the detailed reasoning needed for review.
Sources: [1]
Less reading before the human decision
The evaluation used 183 synthetic patient cases and more than 75,000 trial annotations. Retrieval retained over 90% of relevant trials while selecting under 6% of the initial collection. In a separate user study, assisted screening took 42.6% less time. Those results concern finding and reviewing candidates.
Sources: [1]
A shortlist is not admission to a trial
Missing clinical information and the study team’s final checks still matter. The reported evaluation did not show increased real patient enrolment or better treatment outcomes. Its valuable contribution is a navigable explanation of why a trial may fit, allowing a person or clinician to pursue the next conversation.
The original sources
- US National Library of MedicineAbout TrialGPT
- Nature CommunicationsMatching patients to clinical trials with large language models