NIH's TrialGPT cuts clinical trial patient screening time by 40% using large language models

Finding the right clinical trial for interested patients is a time-consuming and resource-intensive process for clinicians, slowing down important medical research. Clinicians must manually navigate the vast and ever-changing catalog of trials listed on ClinicalTrials.gov, assessing eligibility criteria against individual patient profiles.

NIH researchers from NLM and the National Cancer Institute developed TrialGPT, an LLM-powered framework that processes a patient summary, identifies relevant trials from ClinicalTrials.gov, excludes ineligible ones, and generates an annotated ranked list of trials with eligibility explanations for clinician review.

In a pilot user study, clinicians using TrialGPT spent 40% less time screening patients while maintaining the same level of accuracy as manual review. TrialGPT's criterion-level eligibility predictions achieved nearly the same accuracy as human clinicians when benchmarked against over 1,000 patient-criterion pairs.

• LLMs can match clinician-level accuracy on structured eligibility screening tasks, enabling meaningful time savings without sacrificing precision.
• Ranking and annotating trial matches — rather than just flagging eligibility — is key to making AI-assisted trial matching actionable for clinicians.
• Responsible AI deployment in clinical research requires fairness evaluation across diverse and underrepresented populations before broad rollout.