Abstract
<title>Abstract</title> <p>Background Accurate identification of immune-related adverse events (irAEs) in structured electronic health record (EHR) data remains a major challenge due to the lack of specific diagnostic codes and inconsistent documentation. We propose a clinically grounded, generalizable framework for phenotyping irAEs in structured data, using immune-mediated diarrhea and colitis (IMDC) as a representative use case. Methods Using a large, curated EHR-derived dataset of adults with advanced non-small cell lung cancer (NSCLC) treated with immune checkpoint inhibitors (ICIs), we developed a rule-based phenotyping framework informed by clinical practice guidelines and the clinical course of irAEs. The approach integrates (1) treatment exposure, (2) temporally anchored immunosuppressive therapy, and (3) diagnostic code patterns to identify moderate-to-severe irAEs, with IMDC used to enhance specificity. Results We developed and demonstrated a guideline-informed framework for identifying complex treatment-related toxicities in structured EHR data, using immune-mediated diarrhea and colitis (IMDC) as a representative use case. Among 14,659 eligible patients treated with immune checkpoint inhibitors, 576 (3.93%) met criteria for potential grade 2–4 immune-related adverse events (irAEs). Of these, 58 (0.40%) were further classified as IMDC based on temporally aligned diagnostic codes. The framework demonstrated the feasibility of identifying clinically meaningful adverse events using structured data elements and guideline-informed temporal relationships. Conclusions This study presents a scalable, clinically interpretable, and generalizable framework for phenotyping complex clinical outcomes in structured EHR data without reliance on advanced computational methods. By leveraging treatment patterns, diagnostic codes, and guideline-informed temporal relationships, the approach enhances reproducibility and portability across datasets and healthcare settings. Clinically grounded, rule-based phenotyping may serve as a practical complement to machine learning and natural language processing approaches when developing computable phenotypes for complex treatment-related toxicities.</p>