Volatility-aware multimodal sleep modeling with latent State inference and Pareto-optimized sensor fusion

Traditional sleep monitoring relies on unimodal physiological sensor signals to classify sleep stages and transitions, yet dynamic multi-sensor fusion techniques remain limited. Here, we present a lightweight Volatility-Driven Hidden State Modeling (VDSM) framework that integrates volatility analysis and probabilistic modeling for adaptive sleep decoding to optimize sensor selection and sleep state detection using sleep recordings from 22 participants in a crossover study under placebo and temazepam conditions. Generalized Autoregressive Conditional Heteroskedasticity (GARCH) modeling quantifies intra-stage volatility across temporal and sensor-specific levels and informs Hidden Markov Modeling (HMM) sleep state modeling. Multi-level sensor fusion improves classification accuracy, transition detection and mean dwell time for each sleep stage, hidden state, and sensor configuration. The optimized EEG-EMG-EOG sensor setup achieved the highest metrics (accuracy (92 %), precision (94 %), and F1-score (0.93)) across age groups and drug conditions, significantly outperforming unimodal setups. Furthermore, multimodal fusion improved transition detection sensitivity (up to 98 %) under temazepam, reduced volatility by 35 % across frequencies, and improved sleep stability. Latent state inference enables the identification of probabilistic transitions between hidden sleep stages, improving both transition detection and stability assessment. Pareto multi-objective optimization evaluated trade-offs between detection accuracy and computational efficiency, with the EEG-EMG-EOG configuration achieving the lowest volatility variance (0.07) and highest signal reliability (SNR: 25.6 dB). In addition to improving sleep stage characterization, VDSM optimizes sensor selection by analyzing age related changes in sleep dynamics, assessing the effects of sedatives on sleep stability, improving multi-sensor decision systems for physiological monitoring, and supporting data fusion where sleep instability is present. • Volatility informed features enable hierarchical modeling decoding of latent sleep state transitions. • GARCH models track intra-stage variance in multimodal sleep signals. • HMM-based latent state inference decodes probabilistic sleep state transitions. • Multimodal sensor fusion improves classification and stability metrics. • NSGA-II optimizes trade-offs between accuracy, computational cost, and signal redundancy.