Agentic Pipeline for Large-Scale Clinical Data Queries in Enterprise Electronic Health Records
- 1Boston Children Hospital (Harvard Medical School)
Abstract
Using electronic health record (EHR) data to answer clinical research questions is challenging. It requires understanding the research question, finding relevant information across different sources, and combining that information to produce meaningful results. We developed an agentic large language model (LLM) pipeline that decomposes investigator-authored requests into source-specific tasks, generates and executes SQL on EHR, integrates structured and unstructured EHR evidence, and applies study-specific clinical note adjudication. We evaluated this workflow against data analyst-identified cohorts across 50 pediatric orthopedic clinical data query requests. Of these, 43 were included in manual-review-adjusted performance analyses. The median adjusted positive predictive value, sensitivity among patients identified by either approach, and corresponding F1 score were 0.93 (95% CI, 0.88-0.98), 0.92 (95% CI, 0.90-0.95) and 0.90 (95% CI, 0.87-0.94), respectively. These findings support the feasibility of translating natural-language research intent into executable, autonomous multi-stage EHR workflows.
Related
- Beyond Billing Codes: An Agentic Large Language Model Pipeline for Extracting Glaucoma Procedural Data from Free-Text Clinical Notes
- Expert Evaluation of Clinical AI Tools on Real Point-of-Care Clinical Queries
- An Agentic AI Pipeline for an Evidence-Grounded Environmental Health Public Information System: Development and Evaluation
- Real-world evaluation of a multilingual agentic retrieval-augmented clinical AI framework for evidence-grounded medical decision support
- Multimodal ensemble machine learning pipeline for oocyte fertilization prediction using first polar body morphology and clinical factors
BibTeX
@misc{pawar2026agentic,
title = {{Agentic Pipeline for Large-Scale Clinical Data Queries in Enterprise Electronic Health Records}},
author = {K. Pawar and N. Tavabi and A. Kalepalli and N. Kalepalli and M. Blackmon and A. Williamson and A. Kiapour},
howpublished = {medRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.05.26367378},
url = {https://www.medrxiv.org/content/10.64898/2026.11.05.26367378v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Agentic Pipeline for Large-Scale Clinical Data Queries in Enterprise Electronic Health Records
AU - K. Pawar
AU - N. Tavabi
AU - A. Kalepalli
AU - N. Kalepalli
AU - M. Blackmon
AU - A. Williamson
AU - A. Kiapour
PY - 2026
DA - 2026/11/05
PB - medRxiv
DO - 10.64898/2026.11.05.26367378
UR - https://www.medrxiv.org/content/10.64898/2026.11.05.26367378v1
AB - Using electronic health record (EHR) data to answer clinical research questions is challenging. It requires understanding the research question, finding relevant information across different sources, and combining that information to produce meaningful results. We developed an agentic large language model (LLM) pipeline that decomposes investigator-authored requests into source-specific tasks, generates and executes SQL on EHR, integrates structured and unstructured EHR evidence, and applies study-specific clinical note adjudication. We evaluated this workflow against data analyst-identified cohorts across 50 pediatric orthopedic clinical data query requests. Of these, 43 were included in manual-review-adjusted performance analyses. The median adjusted positive predictive value, sensitivity among patients identified by either approach, and corresponding F1 score were 0.93 (95% CI, 0.88-0.98), 0.92 (95% CI, 0.90-0.95) and 0.90 (95% CI, 0.87-0.94), respectively. These findings support the feasibility of translating natural-language research intent into executable, autonomous multi-stage EHR workflows.
ER -