Privacy-Aware Adaptive Mental State Intelligence for AI-Assisted Decision-Making in Digital Mental Health Systems: Architecture and Empirical Evaluation

Opportunities to anticipate changes in depressive symptoms are provided by longitudinal digital mental-health records, but the usefulness of algorithmic recommendations and the preservation of model utility under privacy protections cannot be established from predictive accuracy alone. A Privacy-Aware Adaptive Mental State Intelligence (PA-AMSI) architecture was specified and evaluated using the Corona Health smartphone dataset. Historical Patient Health Questionnaire-9 (PHQ-9) assessments were used to predict five symptom-severity categories 14–30 days ahead. Following duplicate-origin correction and participant-disjoint partitioning, 6,314 forecasting samples from 903 participants were retained; 5,025 samples from 722 participants were allocated to development, and 1,289 samples from 181 participants were reserved for a single locked-test evaluation. Persistence, centralized classifiers, simulated federated averaging, and participant-clipped Gaussian-noised training were compared. On the locked test, the highest macro-F1 (0.6541) was obtained with persistence, whereas the highest accuracy (0.7168) was obtained with Random Forest. A macro-F1 of 0.2743 was obtained with the Gaussian-noised model, and none of the 46 severe-class assessments was detected. Under a predefined illustrative loss matrix, a lower mean simulated decision cost was obtained with the questionnaire-only policy (0.5531) than with Random Forest-assisted (0.6175) and privacy-noised (0.6594) policies. Higher costs for the AI-assisted policies were supported by participant-cluster bootstrap intervals. It is therefore shown that privacy assumptions, class-specific reliability, and decision consequences must be assessed together before digital mental-health prediction systems are considered for clinical evaluation.