An IoMT-Based Circadian-Mood Digital Twin for Pre-Symptomatic Mental Health Prediction via Federated Learning and Differential Privacy

Continuous mental health monitoring using wearable sensors in the Internet of Medical Things (IoMT) offers a paradigm shift toward pre-symptomatic clinical intervention. However, prevailing predictive models rely on empirical correlation rather than physiological dynamics and face severe data privacy obstacles under healthcare regulations. This paper presents a biologically grounded, privacy-preserving IoMT framework that models patient-specific sleep homeostatic pressure (S) and endogenous circadian drive (C) via the Two-Process Model to construct an individualised Circadian-Mood Digital Twin. Features extracted from 30-day longitudinal records across 10 virtual patients—encompassing heart rate variability, sleep-stage architecture, voice spectral envelopes (13 Mel-Frequency Cepstral Coefficients), and ambient photic exposure—are applied to forecast depressive (PHQ-9) and anxiety (GAD-7) trajectories over 24-h, 36-h, and 48-h horizons. Decentralised model training is executed across five virtual clinical nodes using Federated Averaging (FedAvg) combined with L2 gradient clipping and Gaussian Differential Privacy (DP). Empirical evaluation demonstrates a predictive Mean Absolute Error of 7.891 points for PHQ-9 and 6.071 points for GAD-7 at a 24-h horizon. The early warning scoring mechanism attained an episode detection sensitivity of 90.0% with an average clinical lead time of 71.1 h, while threshold tuning reduced hourly false alarms to 7.2%. Global convergence occurred within 10 communication rounds under a cumulative privacy budget of ε ≈ 96.1. The results validate the feasibility of combining computational chronobiology with privacy-preserving distributed learning for proactive psychiatric monitoring.