Reproduce “Matching Patients to Clinical Trials with Large Language Models” on OpenResearch

About this paper

Qiao Jin, Zifeng Wang, Charalampos S. Floudas, Fangyuan Chen, Changlin Gong, Dara Bracken-Clarke, Elisabetta Xue, Yifan Yang and 2 more

Researchers from NCBI and the University of Illinois Urbana-Champaign developed TrialGPT, an end-to-end framework leveraging large language models to automate patient matching to clinical trials. The system achieved a 42.6% reduction in screening time for medical professionals and demonstrated superior performance compared to baseline models in trial ranking and exclusion tasks.

Key insights

• LLMs can generate high-quality, patient-specific keywords for initial trial retrieval that are more effective than manually generated or raw patient note queries.
• A hybrid retrieval strategy combining lexical and semantic methods with LLM-generated keywords efficiently reduces the candidate trial pool by over 90% while maintaining high recall.
• LLMs are capable of providing accurate, explainable criterion-level eligibility predictions, including the exact supporting sentences from patient notes, which enhances transparency and trust.

Summary by alphaXiv.