AI can now screen patient records for clinical trials faster than humans, but only for complex medical histories

Researchers tested large language models on the task of reading patient records and flagging who qualifies for clinical trials. The best model (MedGemma with a retrieval strategy) scored 89% accuracy, suggesting AI could speed up a notoriously slow bottleneck that causes trials to fail when enrollment stalls.

Clinical trial enrollment is one of the few places where AI has a clear, measurable job: sift through thousands of patient records and flag those who meet eligibility criteria. This is a place where AI could actually ship because the problem is concrete, the stakes are high (failed trials waste months and millions), and success is measurable. The paper shows AI works best on complex eligibility rules that require reasoning across a patient's full medical history, but still only marginally improves performance on simpler checks like lab test values. Real adoption will depend on hospitals and trial sponsors accepting AI-flagged candidates and measuring whether enrollment actually accelerates in practice.