Computable longitudinal patient journeys from structured and unstructured EHR data

Most electronic health record (EHR) data reside in unstructured text, recorded in encounters between health care providers and patients. These encounters capture symptom narratives, adverse effects, treatment rationales, adherence challenges, family history and other nuances that are not coded or only partially reflected in structured fields such as diagnosis and procedures codes. Yet the majority of real-world evidence studies rely on structured data fields. Here we present a novel approach that uses large, pre-trained language models to accurately extract computable clinical data from unstructured EHR text. Physician adjudication confirmed high accuracy against a blinded expert reference standard, with high expert physician inter-reviewer agreement. The framework is disease-agnostic and scales to large patient datasets across all clinical conditions. The extracted clinical entities are integrated with structured EHR data, embedded with medical ontologies, and organized into a knowledge graph (KG), allowing all relationships between variables to be interrogated at the level of individual patients or at scale. An agentic-friendly programmatic interface navigates large-scale KG data, making analyses rapid, easy and cost-effective, both for exploration and hypothesis testing. This human-guided query environment enables acceleration of labor-intensive tasks while study design, analytic decisions and interpretation remain under investigator control. To demonstrate clinical utility, we conducted a large-scale longitudinal analysis of treatment responses among individuals initiating glucagon-like peptide-1 receptor agonists (GLP-1 RAs) therapy. We identify large numbers of patients initiating GLP-1 RAs with high confidence, reconstruct patient-level trajectories, and model longitudinal changes in weight and hemoglobin A1c (HbA1c) following GLP-1 RA initiation, revealing association between these parameters and response to therapy. Further, we characterize outcomes captured only in unstructured clinical notes and show time-to-event outcomes, moving beyond fixed-timepoint endpoints to provide a comprehensive understanding of real-world treatment responses.