Publications

Despite the growing availability of Electronic Health Record (EHR) data, multi-institutional research remains challenging due to differences in data schemas, coding systems, local vocabularies, documentation practices, and limited standardized pre-processing tools. These barriers reduce reproducibility and hinder the development of robust data representations for collaborative translational research. We introduce the Pipeline for Harmonizing Electronic Health Record Data for Translational Research (PEHRT), a transparent end-to-end framework that transforms relational EHR data into analysis-ready resources through standardized pre-processing and representation-based harmonization. Module 1 of the pipeline cleans structured data, processes clinical notes, generates executable cohort definitions from natural language requests, performs transparent code mapping and roll-up, and records all pre-processing decisions in audit logs. Module 2 learns institution-specific embeddings from pointwise mutual information matrices, aligns them across institutions using a summary-level matrix completion algorithm, and integrates them with pre-trained language model embeddings to produce harmonized representations for collaborative analyses. We validate PEHRT through extensive analyses of real-world EHR data, including end-to-end pre-processing, local laboratory code harmonization, embedding quality evaluation across diverse healthcare systems, and downstream feature selection, predictive modeling, and cross-site transportability analyses. Together, these results establish PEHRT as practical infrastructure for reproducible, scalable, and collaborative EHR research.

Introduction: Advanced therapies (AT) are a cornerstone of management of moderate-to-severe Crohn's disease (CD) and ulcerative colitis (UC). While endoscopic healing (EH) is recognized as a therapeutic target, there is limited data on long-term outcome of EH, in part due to paucity of efficient methods to identify EH in large datasets.

Methods: Using a multi-institution electronic health record (EHR) cohort, we developed and validated a large language model (LLM) based agent to identify severity of endoscopic activity on text reports from colonoscopies that were prospective scored using SES-CD (CD) or modified Barron's score (UC). The primary cohort for the analysis consisted of patients with CD or UC initiating advanced therapy (AT) for their IBD between 1998–2025. We identified all patients who had a colonoscopy between 6–18 months after AT initiation. This endoscopy report was scored using the LLM agent and classified as severe, moderate, mild, or endoscopically healed (remission). We examined the risk of IBD-related surgery, hospitalization, and corticosteroid use at 2 years, 5 years, and till end of follow-up after this index colonoscopy using adjusted Cox-proportional hazards models.

Results: Using a training dataset of 672 patients with CD and 293 with UC, we were able to develop an LLM agent that defined remission with an accuracy of 84–88% for CD and 81–93% for UC. This was independently validated in two additional datasets. The analysis of EH on long-term outcomes including 1,944 patients with CD and 1,399 with UC with the most-common AT being TNF-antagonists. In CD, compared to patients with severe endoscopic activity at follow-up colonoscopy, patients who had moderate (HR 0.52, 95% CI 0.32–0.84), mild (HR 0.31, 95% CI 0.18–0.52) or endoscopic remission (OR 0.14, 95% CI 0.08–0.25) had lower likelihood of surgery after 2 years of follow-up from the colonoscopy (p < 0.001). EH was also associated with reduced risk of IBD-related hospitalizations, stricturing/penetrating complications, and corticosteroid use. In UC, compared to those with severe endoscopic activity, the presence of mild (HR 0.34, 95% CI 0.13–0.94) disease activity or endoscopic healing (HR 0.11, 95% CI 0.03–0.40) was associated with a reduced risk of resection at 2 years following endoscopic assessment. EH was similarly associated with reduced need for IBD-related hospitalization.

Conclusions: LLM agents are able to accurately infer endoscopic activity from text reports in multi-institution EHR datasets. Endoscopic healing following initiation of AT is associated with superior long-term outcomes in both CD and UC.