Physiological-Interaction-Aware Net for Automatic Sleep Staging from Concurrent Spatio-Time-Frequency View

OBJECTIVE: Accurately sleep staging is crucial for sleep monitoring and disorder diagnosis. Physiological systems, such as the brain, heart, and respiratory system, exhibit distinct interaction patterns at different stages of sleep, and these interactions are reflected through signal coupling mechanisms. The physiological interactions assist in identifying and classifying sleep stages, while also supporting the maintenance of physiological balance. However, the present researches seldom consider the physiological interactions in sleep, thus limiting the classification accuracy. METHODS: This paper develops physiological-interaction-aware net (PIANet) for automatic sleep staging from concurrent spatio-time-frequency view. PIANet consists of five modules, i.e., functional connectivity representation, spatio-temporal feature extraction, time-frequency representation, time-frequency feature extraction, and cross-view feature fusion. The model integrates both Euclidean and non-Euclidean spaces to capture the complex dependencies among multiple physiological systems and signals during sleep. By incorporating both spatiotemporal and time-frequency features, the model effectively identifies the physiological interactions across different sleep stages. RESULTS/CONCLUSION: Experimental results demonstrate that the model outperforms existing methods on the ISRUC-S1, ISRUC-S3, and DOD-H datasets, achieving classification accuracies of 0.825, 0.836, and 0.891, and F1 scores of 0.814, 0.826, and 0.839, respectively.