TriNetX adds human-validated facts from clinical notes for researchers

A larger source of clinical information

The company’s natural language processing system extracts information from more than 800 million clinical notes and reports. TriNetX expects that base to reach approximately one billion by the end of 2026.

Clinical notes can contain details that are often absent from structured medical records, including disease stage, general condition and biomarker results. These details can help researchers expand patient cohorts and build more detailed clinical phenotypes, but extracting them consistently from free text is difficult.

TriNetX says its system combines artificial intelligence with clinical and computer science expertise. The pipelines are trained on a large collection of annotated reference notes and include human intervention. Extracted facts are connected to structured electronic medical record data for the same de-identified patients.

Three checks support each extracted fact

TriNetX describes three safeguards for the data. First, clinical experts review and correct the information labelled by the AI, and the models are retrained using those corrections. Each module must reach a precision of at least 0.90 against notes annotated by humans.

Second, the company carries out clinical confirmation by testing extracted facts against expected clinical patterns, rather than relying only on model indicators. TriNetX gives Eastern Cooperative Oncology Group performance scores as an example: survival curves showed lower survival among patients with a score of 2 or higher.

Third, the company says it continuously monitors model performance as more healthcare organisations join its clinical notes programme and as new types of notes are added. The models are refined when variations appear.

Early findings and current coverage

According to TriNetX, information extracted from notes increased coverage of smoking status from 22% to 67%. In a lung cancer cohort study, the note-based information supported six times more patients than structured data alone.

The company also says it identified ECOG and Karnofsky performance scores for 292,000 oncology patients, although none of that information appeared in the structured fields.

The available areas currently include social history, such as smoking and alcohol use, cardiology, pulmonology, general oncology and biomarkers, and liver disease, among others. The extracted clinical note data are available to TriNetX network members through TriNetX LIVE and downloadable datasets.

Why validation matters for research

TriNetX’s announcement addresses a central question for researchers using AI to analyse clinical notes: whether extracted information can be trusted in real-world research. The company says its process goes beyond extraction by measuring performance against human-annotated notes, attaching a confidence score to each fact and checking the results against clinical expectations.

TriNetX CEO August Calhoun said the company’s approach combines scientific rigour and human expertise within a global federated network, where data remain with the organisations that generated them. The company’s medical informatics director, Matvey Palchuk, said that no system can extract every detail from clinical notes with both precision and relevance because of the nature of clinical language. He said validation is therefore important to extracting facts that can support evidence production.

For researchers, the offering provides a way to combine information from narrative notes with structured records while retaining a confidence measure and human review process. TriNetX says the capability is now accessible to members of its network.