Real-time sleep disorder monitoring design using dynamic temporal graphs with facial and acoustic feature fusion

Abstract

Introduction: Sleep disorders pose significant risks to patient safety, yet traditional polysomnography imposes substantial discomfort and laboratory constraints. We developed a non-invasive multimodal monitoring system for real-time sleep pathology detection. Methods: We integrated facial expression analysis via deep convolutional neural networks with audio signal processing for breathing pattern detection. Heterogeneous data streams were unified into dynamic graph representations, with graph neural networks modeling spatiotemporal patterns of sleep pathologies. Results: The system accurately detected sleep apnea, restless leg syndrome, and cardiovascular irregularities with 10.7-s average delay and 94.6% clinical agreement, achieving diagnostic accuracy comparable to polysomnography. Conclusion: This framework enables continuous non-invasive monitoring for point-of-care screening and home-based management, potentially expanding sleep medicine access for underserved populations.