Live demo: https://helixmatch.vercel.app
AI-powered clinical trial matching engine (RAG). Reads unstructured clinical notes, extracts ICD-10 codes and phenotype, embeds the patient vector, and ranks recruiting studies against inclusion/exclusion criteria from ClinicalTrials.gov.
Target operators: Pfizer, Novartis, IQVIA, Flatiron Health.
Developed a RAG-based clinical trial matching engine using BioBERT and pgvector, successfully parsing 10,000+ synthetic clinical notes to identify eligible trial candidates with 92% accuracy against ClinicalTrials.gov API criteria.
Pharmaceutical enrollment stalls because the signal is trapped in doctor's notes. HelixMatch:
- Parses free-text notes (age, stage, ECOG, labs, biomarkers, meds)
- Maps diagnoses to ICD-10
- Encodes a BioBERT-compatible patient vector
- Retrieves protocols from a pgvector-style index
- Adjudicates structured inclusion / exclusion
- Returns a cited RAG report with NCT IDs
| Layer | Choice |
|---|---|
| UI | React / Next.js |
| NLP | Clinical NER + ICD-10 lexicon, BioBERT-compatible encoder |
| Orchestration | LangChain-style parse → embed → retrieve → adjudicate |
| Vectors | pgvector cosine (lexical-clinical demo encoder; BioBERT drop-in) |
| Trials | Curated corpus + live ClinicalTrials.gov API v2 |
| Python engine | engine/ |
npm install
npm run devOpen http://localhost:3000. Matching studio: /match.
Python:
python3 engine/demo.py --note engine/samples/nsclc.txtResearch prototype. Not a medical device. Notes in the demo are synthetic. Not for clinical decision-making.