Project 6: Clinical Trial Matcher¶
Mentors: To be announced
Problem: Real, currently-recruiting clinical trials are publicly searchable but written in legal-style eligibility text a patient can't quickly match against their own situation.
Context: Pulls live data from the ClinicalTrials.gov API (public, no key required), ranking with off-the-shelf open-source sentence-transformers embeddings. No training, deliberately absent so students feel the difference between RAG-style ranking and an actually fine-tuned model, as in Projects 3 and 5.
Goals: Given a plain-language description of a patient's situation, surface and rank real, currently-recruiting trials, and summarize each one's eligibility criteria so a human doesn't have to parse legal-style trial text.
Deliverables: A free-text input + ranked-trial-list app built from (1) a ClinicalTrials.gov query for a broad condition match, (2) sentence-transformers embeddings of the query and each trial summary, ranked by cosine similarity, and (3) an LLM call per top-ranked trial to compress its eligibility criteria into a few plain-language bullets. The most "full pipeline" project before the capstone: API + embeddings + LLM, no training.
Showcase: TBD
References:
- Matching Patients to Clinical Trials with Large Language Models (TrialGPT, Nature Communications, 2024): NIH/NLM's own retrieve-then-explain pipeline for patient-to-trial matching, the same shape as this project's embed-rank-summarize approach.
- TrialGPT source code: the open-source implementation from NLM's intramural research program, useful as a reference for how to structure the retrieval and eligibility-summarization steps.