Maintaining EHR-based Predictive Clinical Decision Support in Mental Health against Temporal Shifts

Electronic health record (EHR)-based risk prediction offers opportunities for early intervention in mental health, but models may remain in use as patient populations, care practices, and documentation change. Such temporal shifts can reduce predictive performance, while updating models from recent data alone may be unstable when adverse outcomes are uncommon. We developed Adversarial Drift-Aware Predictive Transfer (ADAPT), a distributionally robust approach to improve the temporal durability of mental health risk prediction. ADAPT uses variation in models estimated from previous calendar periods to characterize plausible changes in predictor-outcome relationships. Data from the current period constrain which historical patterns remain relevant, allowing the method to seek robustness against subsequent shifts while preserving contemporary discrimination. We retrospectively evaluated ADAPT using EHRs from 889,445 patients with mental health diagnoses at Mass General Brigham and Duke University Health System, covering one-year risk prediction for 28 outcomes. In the worst 20% of year-pair comparisons, ADAPT improved AUC over all comparators by 0.016-0.036 at Mass General Brigham, 0.011-0.018 at Duke, and 0.008-0.037 after transport. For suicide-attempt prediction, contemporary AUC at Mass General Brigham was 0.82 compared with 0.80 for the best comparator. Eight-year relative AUC degradation was 7.7%, compared with 12.1% for pooled-data training. These results suggest that information from prior temporal shifts can improve robustness to future changes without sacrificing predictive performance for current patients.