Leveraging Electronic Health Record Data and Artificial Intelligence to Develop a Crosswalk Tool for Personalized Clinical Experience Profiles of Emergency Medicine Residents

PROBLEM: Graduate medical education requires learners to acquire broad clinical exposures to meet core competencies for unsupervised practice, but variability in clinical learning environments and reliance on resource-intensive assessments hinder precise assessment of trainees' clinical experiences. Electronic health records hold promise for precision medical education, yet manual mapping of International Classification of Diseases, Tenth Revision (ICD-10) codes to specialty-specific clinical practice domains limits scalability. APPROACH: The authors leveraged electronic health record data and artificial intelligence (AI) to map residents' encounter diagnoses to the American Board of Emergency Medicine's Model of the Clinical Practice of Emergency Medicine (MCPEM). Resident encounters across 3 sites at a single academic system (January 1 to October 31, 2023) were analyzed with an AI model, mapped to MCPEM categories with ICD-10 descriptors, and quantified with vectors to match to the closest MCPEM category. Faculty raters validated the most common mappings iteratively, which were subsequently integrated into interactive learner dashboards. OUTCOMES: Among 119,320 encounters, 5,960 unique ICD-10 descriptors (1,126 stem codes) were identified. For the 650 most common diagnoses, 507 (78.0%) of emergency department diagnosis text descriptors were determined as valid mappings to an MCPEM subcategory. In mappings where faculty were discordant with the lowest distance mapping, 171 of 305 alternative subcategory mappings (56.0%) achieved agreement, increasing the concordance between reviewers to 515 of 650 (79.2%) overall. Interactive dashboards displayed resident-level case mix mapped to MCPEM categories, with anonymized peer comparisons and program-level aggregates, enabling identification of patterns and gaps by domain. NEXT STEPS: Planned work includes iterating AI-automated mappings by expanding inputs beyond diagnoses, engaging wider stakeholder review of mapping validations, and assessing generalizability to other specialties' content outlines to produce a scalable and reproducible model to increase the precision of feedback loops to inform graduate medical education, the clinical learning environment, and training design.