Calibrated and Explainable Temporal Forecasting of Psychoemotional Risk from Longitudinal Mobile-Health Data

Mobile-health systems generate repeated, time-stamped observations, but future-risk forecasts must be reliable rather than merely discriminative. We evaluate Temporal Calibrated Explainable Risk Forecasting (T-CERF) for irregular ecological momentary assessment histories. Study A used 5726 forecasting examples from 542 Corona Health participants, a 56-day history window, and a 28-day primary horizon. On the fixed participant-independent test set, T-CERF achieved a PR-AUC of 0.867, ROC-AUC of 0.946, and Brier score of 0.083. Random Forest matched its ROC-AUC (0.947 versus 0.946), while the Persistence baseline yielded a higher F1-score (0.786 versus 0.761). The PR-AUC difference from the matched current-only model was +0.023, but its paired participant-level bootstrap interval included zero. Across five additional participant-level splits, ΔPR-AUC averaged +0.0106 (SD 0.0262; range −0.0298 to +0.0399), with one split favoring the current-only model. Thus, a consistent discriminative benefit from temporal history was not established. Temperature Scaling improved calibration, while ensemble uncertainty supported selective prediction. A separate HBSC Kazakhstan analysis remains exploratory. T-CERF is therefore presented as a calibrated, uncertainty-aware mobile-health forecasting framework rather than as a hardware or network-level IoT implementation.