An informatics framework to harmonize electronic health record medication data for managed care analytics and artificial intelligence applications

BACKGROUND: Artificial intelligence (AI) applications in managed care pharmacy depend on semantically consistent medication data, yet heterogeneous medication identifiers across real-world electronic health records (EHRs) could undermine analytic fidelity and risk propagating classification errors. To enable transportable, reproducible AI tools, methods for harmonizing disparate medication identifiers (eg, National Drug Code [NDC] and Multum drug synonym ID) to standardized vocabularies are required. OBJECTIVE: To develop and evaluate an informatics framework that harmonizes heterogeneous medication identifiers into standardized ingredient- and pharmacologic-class representations, establishing the semantic integrity required to support transparent, reliable downstream managed care analytics and AI developments. METHODS: We conducted a retrospective analysis of discharge medication records, excluding health supplements and nonmedicated agents, from EHRs of adults aged 65 years or older who had at least 1 hospitalization between 2020 and 2024 at a tertiary medical center (Buffalo General Medical Center). Heterogeneous identifiers (Multum, NDC, RxCUI) were harmonized using a 2-layered architecture anchored to RxCUI ingredient [IN] concepts and abstracted to Anatomical Therapeutic Chemical (ATC) classification. Deterministic crosswalks were derived from the RxNorm Full Monthly Release and complemented with a validated TriNetX reference file. Unmapped records underwent structured string-based reconciliation through alignment of medication name strings between raw records and reference terminology sources. A pharmacy expert manually validated the string-based reconciliation process and the RxCUI [IN]-to-ATC mapping. Framework feasibility was assessed by calculating mapping success rates and evaluating reconciliation and correction metrics. RESULTS: Analysis of 214,080 records revealed significant terminology heterogeneity, with 53.0% Multum Drug Synonym IDs as nonstandardized identifiers requiring string-based reconciliation. Mapping using deterministic crosswalks achieved a 100% initial mapping rate, with 30% to 35% of unique records requiring ATC assignment corrections. While string-based alignment achieved initial match rates of 78.5% and 75.5% for RxCUI [IN] and ATC codes, respectively, reconciliation required corrections for 57.4% of assignments. Primary drivers for corrections included crosswalk omissions for branded formulations, lexical misclassifications, and taxonomic ambiguities that required clinical context for accurate ATC assignment. Examples of taxonomic ambiguities include variations in ATC classification based on indication or route of administration (eg, moxifloxacin may be mapped to J01MA14 or S01AE07). CONCLUSIONS: This informatics framework establishes the architectural foundation necessary to transform raw EHR medication records into standardized representations at the ingredient and pharmacological levels, providing a critical semantic foundation for transparent, transportable, and reproducible AI applications in managed care.