Back to Bigger Picture

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.

AI-generated illustration: Patient information is compared with individual trial eligibility criteria.
Patient information is compared with individual trial eligibility criteria.AI-generated illustration of the idea, not a photograph of the actual people, equipment or event.
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

  1. Retrieve possible trials

    Patient information narrows a large trial collection.

  2. Check each criterion

    The language model compares eligibility rules with the case.

  3. 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.

Sources: [1][2]

The original sources

  1. US National Library of MedicineAbout TrialGPT
  2. Nature CommunicationsMatching patients to clinical trials with large language models

What to explore next

Can a hospital spot an empty appointment before it happens?Counting fish while they move along the sorting belt
Explore all 99 applications