Medication-Wide Association Study of Alzheimer's Disease and Related Dementias: Identifying Drug Candidates from Electronic Health Records through Explainable AI.

Objective: Alzheimer's disease (AD) is a leading cause of death and disability, and treatment options for Alzheimer's disease and related dementias (ADRD) remain limited. We applied a data-driven, mechanism-agnostic Medication-Wide Association Study Plus (MWAS+) framework to identify candidate medications associated with ADRD using longitudinal electronic health record data and explainable artificial intelligence (AI). Methods: We used Veterans Health Administration electronic health record data from January 1999 to May 2022. The initial study population comprised 8,424,715 Veterans aged 65 years or older. Cases were defined by ADRD-related diagnosis codes or ADRD-related medication prescriptions, and controls were free of ADRD diagnosis and ADRD-related medication use. After exclusions and matching on sex, race, age at first encounter, and duration of follow-up, the primary analytic cohort included 505,817 matched case-control pairs (1:1; 1,011,634 Veterans). Longitudinal features were extracted from historical data up to 1 year before the index date and aggregated into 1-year intervals. We developed an upgraded Hybrid Value-Aware Transformer (HVAT 2.0) to jointly learn from longitudinal and nonlongitudinal clinical data while incorporating numerical values associated with clinical concepts, including cumulative medication dose. To enhance interpretability, we applied a medication-specific impact score method to estimate model-derived associations between medication exposure and ADRD risk. Findings: The model demonstrated stable performance across data partitions, with area under the receiver operating characteristic curve values of 0.791 in the training set, 0.772 in the validation set, and 0.775 in the testing set. Metolazone and varenicline were identified as the top 2 candidate medications with negative impact scores, suggesting potentially protective associations with new-onset ADRD. The impact score was -0.196 per unit of cumulative dose for metolazone (1800 mg) and -0.134 per unit for varenicline (280 mg). Although individual-level impact scores varied, most exposed patients had negative scores, including 12,020 of 12,480 metolazone users (96%) and 8,341 of 8,786 varenicline users (95%). Implications: This study demonstrates the feasibility of combining a medication-wide association framework, longitudinal dose-aware modeling, and explainable AI to identify candidate medications for ADRD from real-world electronic health record data. The findings should be interpreted as signals for hypothesis generation rather than evidence of causality. This framework may support prioritization of repurposing candidates for expert review, follow-up cohort validation, and future clinical investigation.