Why the richest clinical detail never reaches research
Electronic health records contain far more information than their coded fields can hold. As the authors of a newly published Nature Medicine paper point out, the majority of EHR data resides in unstructured text — the narrative recorded during encounters between health care providers and their patients. Those encounters capture symptom narratives, adverse effects, treatment rationales, adherence challenges, family history and other nuances that are either not coded at all or only partially reflected in structured fields such as diagnosis and procedure codes.
The result is a persistent blind spot in real-world evidence. Most studies that attempt to learn from routine care still lean on structured data fields, simply because those are the parts of the record that can be queried computationally at scale. The new work proposes an approach intended to close that gap.
A pipeline that makes clinical text computable
The researchers present a novel method that uses large, pre-trained language models to accurately extract computable clinical data from unstructured EHR text. Crucially, extraction is treated as the opening stage of a longer pipeline rather than an end in itself, with the goal of making entire patient histories machine-readable from end to end.
From raw notes to an integrated knowledge graph
After clinical entities are pulled out of free text, they are integrated with structured EHR data, embedded with medical ontologies, and organized into a knowledge graph. That architecture matters, because it preserves the relationships between variables instead of flattening them into disconnected fields.
- Unstructured note text is mined for clinical entities using large pre-trained language models.
- Extracted entities are merged with the structured records already held on the same patient.
- Medical ontologies supply the semantic scaffolding that keeps extracted terms comparable and consistent.
- The knowledge graph links variables so that any relationship between them can be interrogated.
Because the graph is organized at the level of individual patients as well as at population scale, an investigator can trace a single person's journey or examine patterns across an entire cohort using the same underlying representation.
Physician review backs the accuracy claims
Accuracy was not left to be taken on trust. The team submitted its outputs to physician adjudication, which confirmed high accuracy against a blinded expert reference standard. Expert physician reviewers also demonstrated high inter-reviewer agreement, an important signal about how consistently the extracted data could be judged against that standard.
The authors are equally explicit about breadth. The framework is described as disease-agnostic, and it is designed to scale to large patient datasets across all clinical conditions rather than being tailored to a single specialty or indication.
Querying at scale while investigators stay in control
To make the knowledge graph usable, the researchers built an agentic-friendly programmatic interface that navigates large-scale graph data. The interface is intended to make analyses rapid, easy and cost-effective, whether the task is open-ended exploration or formal hypothesis testing.
That human-guided query environment is framed as a way to accelerate labor-intensive work rather than to replace scientific judgement. Study design, analytic decisions and interpretation remain under investigator control. In practice, the division of labor is clear: automation handles retrieval and assembly of the data, while researchers retain authority over what questions are asked and what the answers mean.
Demonstrating utility with GLP-1 receptor agonists
To show that the framework has clinical value, the authors conducted a large-scale longitudinal analysis of treatment responses among individuals initiating glucagon-like peptide-1 receptor agonist (GLP-1 RA) therapy. Using their approach, they identified large numbers of patients starting these medications with high confidence, reconstructed patient-level trajectories, and modeled longitudinal changes in weight and hemoglobin A1c (HbA1c) following the start of treatment.
The exercise illustrates the kind of question the pipeline is built to answer. GLP-1 RA use generates a long trail of clinical documentation — eligibility discussions, titration decisions, tolerability issues, adherence problems and follow-up measurements — much of which lives in narrative notes rather than in tidy coded fields. Reconstructing what actually happened to each patient, and then aggregating those individual stories into population-level signal, is exactly the task that unstructured text has traditionally made difficult.
Who is behind the work
The paper is the product of a multidisciplinary author group: Edward Kim, Richard Foty, Avnesh S. Thakor, Jay S. Skyler, Lucy F. Robinson, James D. Park, Monica Kraft, Charles B. Cairns and Vicki Seyfert-Margolis. It was published online on 10 September 2026 in Nature Medicine as an open access article.
Early access and what remains provisional
Nature Medicine has shared the paper ahead of the final version in order to provide faster access to peer-reviewed, accepted research. The article is citable and carries a permanent DOI. Readers should nonetheless note that this version is subject to further edits and will be replaced automatically by the final Version of Record, with all legal disclaimers applying.
Even in its accepted form, the contribution is best understood as infrastructure as much as a single study. The authors argue that their combination of language-model extraction, ontology embedding and knowledge-graph organization addresses a structural problem in clinical research — the mismatch between where clinical detail actually lives and where analysis can reach. By converting encounter narratives into computable entities that sit alongside structured fields, and by exposing the result through a programmatic interface, the framework aims to shorten the distance between a raw note written in a clinic and a testable research question.
Whether that promise holds across the full range of clinical conditions, note styles and health systems is a question that the described disease-agnostic design is meant to invite rather than settle. What the authors offer is a reproducible path: validate extraction against blinded expert review, anchor extracted concepts in medical ontologies, connect everything in a graph, and let investigators query it without surrendering control of the science.
This article is based on reporting by Nature Medicine. Read the original article.
Originally published on nature.com
