MVFSleepNet: Multi-View Fusion Network for Automatic Sleep Staging

The classification of sleep stages holds paramount importance in the field of sleep medicine, particularly for the assessment of sleep quality and diagnosis of sleep disorder. Many deep learning methods for automatic sleep staging rely on the polysomnography (PSG) signals. This work proposes MVFSleepNet, a multi-view fusion network for automatic sleep stage classification. To capture the relationship representation among multichannel PSG signals, we construct two views of raw signal and time-frequency images. For the raw signal input, we utilize the combination of multi-scale 1D-CNN and Structured State Space Sequence model(S4) to capture the temporal features. For the time-frequency input, we employ the combination of S4 and Transformer to learn the spectral representation. To fuse spectral and temporal features, we propose a multi-view attention module to adaptively capture information from two views. Experiments demonstrate that out model achieves superior classification performance compared to state-of-the-art methods. The overall accuracy and macro F1-score (Sleep-EDF-20: 89.5%-83.3%, Sleep-EDF-78: 87.0%-81.5%) observed across two public datasets indicate that the proposed model effectively extracts and integrates information from two views, which contributes to precise sleep staging in clinical applications.