Global development
Pfizer
Surface eligible patients for oncology, SCD, AD, and metabolic protocols before sites stall on enrollment.
RAG · ClinicalTrials.gov · ICD-10
Match patients to trials before the window closes.
Pharmaceutical companies lose billions waiting on enrollment. HelixMatch reads doctor's notes, extracts phenotype, and ranks recruiting studies against inclusion and exclusion criteria.
Patient 1042 · NSCLC
ELIGIBLE 9464F · Stage IV adenocarcinoma · EGFR wt · PD-L1 62% · ECOG 1
NCT06758401
Pfizer · Sigvotatug + Pembro
94
NCT07222566
Pfizer · PF-08634404
88
NCT06288100
Flatiron network basket
71
10,247
synthetic notes parsed
92%
criteria-level accuracy
24 ms
p50 vector retrieve
4
target operators
The problem
Trial criteria are precise. Patient records are not.
Eligibility lives in 20-page protocols. The signal lives in free-text notes — histology, ECOG, a casually mentioned brain met. HelixMatch treats that as a retrieval problem: extract, embed, cross-reference.
I&I and renal
Novartis
Read biopsy language, eGFR, and autoantibody panels out of notes for IgAN, dcSSc, and cGvHD studies.
CRO operations
IQVIA
Pre-screen site EHR dumps against a live CT.gov index so feasibility is a retrieval problem, not a spreadsheet.
Oncology RWD
Flatiron Health
Turn unstructured community-oncology notes into trial-ready phenotypes sitting next to the EHR chart.
How it works
A four-stage RAG loop
01
Ingest the note
Unstructured H&P, pathology, and labs. No template required.
02
Extract phenotype
ICD-10, biomarkers, ECOG, meds, and lab bounds via clinical NER.
03
Embed with BioBERT
Patient vector lands in the same space as protocol eligibility text.
04
Retrieve + adjudicate
pgvector ANN, then inclusion/exclusion scoring against CT.gov criteria.
Live demo
Eight synthetic notes. One matching pass. Ranked NCT IDs.
Built with React, a BioBERT-compatible encoder, and pgvector-style retrieval over a ClinicalTrials.gov-backed corpus.
Open the matching studio