TriNetX Adds Human-Validated Facts Extracted From Clinical Notes
TriNetX has launched NLP-extracted Facts from Clinical Notes, a capability that gives researchers confidence-scored data points from clinical notes and links them to structured EHR data for the same de-identified patients. The company says the feature is available in the TriNetX LIVE platform and in downloadable datasets.
- TriNetX launched NLP-extracted Facts from Clinical Notes.
- The data comes from more than 800 million clinical notes and reports.
- Each fact receives a confidence score and is tested against human-annotated notes.
- Smoking status coverage increased from 22% to 67%, according to TriNetX.
- The capability is available to TriNetX network members.
From clinical notes to research data
The capability draws on more than 800 million clinical notes and reports, a base TriNetX expects to reach about 1 billion by the end of 2026. The company says the notes contain information that can complement structured records, including disease stage, performance status and biomarker results.
TriNetX said these details can be difficult to extract reliably. Its approach uses natural language processing alongside clinical and informatics expertise, with pipelines trained on a large collection of annotated gold standard notes. The resulting facts are designed to help researchers expand cohorts and create more detailed clinical phenotypes.
The extracted information currently covers social history, including smoking and alcohol status, as well as cardiology, pulmonology, general oncology and biomarkers, liver disease and other areas. The capability is available to TriNetX network members.
Three safeguards for extracted facts
TriNetX said every fact is supported by three safeguards. First, clinical experts review and correct what the AI identifies, and the models are retrained using those corrections. Each module must achieve a precision of 0.90 or higher against human-annotated notes.
Second, the company tests extracted facts against expected clinical patterns, rather than relying only on model metrics. TriNetX cited survival curves based on extracted Eastern Cooperative Oncology Group, or ECOG, performance status as an example. Those curves showed worse survival for patients scoring 2 and higher.
Third, TriNetX continuously monitors model performance as additional healthcare organizations join its clinical notes program and as new note types become available. The company says this process allows the models to be refined as more variation appears in the data.
Every fact also receives a confidence score and is linked to structured electronic health record data for the same de-identified patients. TriNetX said this combination is intended to support research that uses both sources of information.
Early results and implications for researchers
According to TriNetX, notes-derived facts increased smoking status coverage from 22% to 67%. In a lung cancer cohort study, the data delivered 6x more patients than structured data alone. The company also said it surfaced ECOG and Karnofsky performance status for 292,000 oncology patients, while those measures appeared in none of the structured fields.
August Calhoun, CEO of TriNetX, said the company’s distinction lies in combining AI and human expertise across a federated network where data remain within the institutions that generated them. Matvey Palchuk, MD, MS, FAMIA, TriNetX’s chief medical informatics officer, said no system can extract every detail from clinical notes in a way that is both accurate and meaningful, making validation important.
For TriNetX network members, the launch provides access to additional analysis-ready information that may be absent from structured EHR fields. The company’s stated focus is on extracting facts that can be trusted for evidence generation, while its validation and monitoring processes address the limits of automated clinical-language analysis.
