A Multimodal Signal Fusion and Heterogeneous Graph Neural Network-Based Algorithm for Sleep Stage Classification

Sleep staging serves as a critical foundation for deciphering sleep mechanisms and clinical diagnosis, while traditional methods relying on manual feature extraction remain inefficient. Current deep learning models still face three major challenges: insufficient collaborative modeling of multimodal signals, high-frequency information loss, and over-smoothing in graph networks. This study proposes two innovations: 1) A Multichannel Biosignal Fusion Model (MC-GCN) employing a dual-path architecture integrating 3D convolutions and graph convolutions to simultaneously capture time-frequency characteristics and cross-modal correlations; 2) An enhanced Multi-relational Heterogeneous Graph Neural Network (M-HGNN) incorporating a high-frequency separation algorithm for personalized frequency band fusion, combined with a Markov enhancement mechanism to dynamically optimize neighborhood information aggregation. Experimental evaluations on the ISRUC dataset demonstrate that our model achieves an overall accuracy of 0.846 and an F1-score of 0.848, showing significant improvements over baseline models: 15.2% enhancement in multimodal interaction capability, 22.8% increase in high-frequency utilization rate, and 18.6% elevation in high-order information capture efficiency. This work establishes a novel paradigm for intelligent sleep monitoring systems.