A Responsible and Risk-Aware Artifical Intelligence Framework for Multimodal Sleep Cycle Monitoring Using Wearable and Contactless Physiological Sensing

Abstract

The use of artificial intelligence has made persistent and tailored monitoring of physiological functions possible, though with the problem of reliability and safety in deployment . The aspect of sleep monitoring is the most sensitive since it is closely linked with cardiovascular, metabolic, and cognitive well-being. Most current automated systems are based on a small number of sensing modalities and non-transparent learning models which limit robustness and interpretability in real-word settings.

This paper introduces a multimodal artificial intelligence sleep monitoring system that combines heterogenous physiological and behavioral indicators of sleep cycles in a late-fusion learning system. The framework maintains modality- specific representations and enables adaptive temporal integration to enhance resilience in unconstrained conditions. In order to raise its reliability, the system integrates uncertainty-aware inference and intrinsic explainability, enabling predictive confidence estimation and features contributions supplied by individual modalities.

Instead of creating an autonomous diagnostic system, the offered method acts under the version of the decision-support tool because it reveals interpretable outputs and highlights the predictions of low level of confidence to be reviewed by experts. Experimental findings show better performance compared to unimodal baselines especially in physiologically ambiguous sleep states, while maintaining transparency and reliability.

This framework adds a deployable and interpretable solution of intelligent sleep monitoring solution to healthcare.