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

ADAPT improves temporal durability of EHR-based mental health risk prediction, raising AUC in the worst 20% of year-pair comparisons by 0.016–0.036 at Mass General Brigham.

• Why it matters: Mental health risk models can lose accuracy as patient populations, care practices, and documentation shift over time, and retraining on recent data alone can be unstable when outcomes are rare.
• What they did: The study retrospectively evaluated ADAPT on EHRs from 889,445 patients with mental health diagnoses at Mass General Brigham and Duke, predicting 28 one-year outcomes using adversarial drift-aware transfer across calendar periods.
• The result: In the worst 20% of year-pair comparisons, ADAPT improved AUC versus all comparators at both sites and after transport, and for suicide-attempt prediction achieved contemporary AUC 0.82 vs 0.80 for the best comparator.

AI summary of the abstract · may contain errors · not medical advice · read the paper.