Agentic Pipeline for Large-Scale Clinical Data Queries in Enterprise Electronic Health Records

An agentic LLM pipeline for enterprise EHR queries achieved median adjusted positive predictive value 0.93, sensitivity 0.92, and F1 0.90 across 50 pediatric orthopedic requests.

• Why it matters: EHR research questions are hard because investigators must interpret intent, locate relevant data across sources, and reconcile structured and clinical-note evidence.
• What they did: The pipeline decomposes natural-language requests into source-specific tasks, generates and runs SQL, integrates structured and unstructured EHR evidence, and performs study-specific note adjudication.
• The result: Compared with data analyst–identified cohorts, 43 requests supported manual-review-adjusted analyses, showing strong agreement that natural-language intent can be translated into autonomous multi-stage EHR workflows.

AI summary of the abstract · may contain errors · not medical advice · read the paper.