GateSleepNet: A Dual-Level Spatiotemporal Graph-Transformer Architecture for Automatic Sleep Staging
Yikun Feng, Yu An, Hongqiang Sun, Zhihong Li, Xi Zhang · IEEE Transactions on Automation Science and Engineering · 2025
Automatic sleep staging is critical for understanding sleep patterns and diagnosing sleep-related disorders, yet traditional manual scoring methods remain time-consuming, laborintensive, and subjective. Existing research often fails to fully exploit both cross-spatial and local-global temporal information in polysomnography (PSG) data. To address these challenges, we propose GateSleepNet, a novel dual-level spatiotemporal framework designed to handle the unique properties of PSG data, which comprises multi-channel physiological signals such as EEG, EOG, and EMG that capture temporal and cross-spatial interactions. GateSleepNet combines a Global Spatial Encoder and a Temporal Vision Transformer (ViT) Encoder to effectively capture both local and global temporal features. To enhance cross-spatial understanding, the framework incorporates a Global Spatial Encoder that models inter-channel cross-spatial relationships using sparsely connected graphs. A key innovation of GateSleepNet is the dual-level surrogate loss, which combines global epoch-level accuracy with local temporal consistency, ensuring alignment between predictions and actual sleep patterns. Experimental results on PSG datasets demonstrate the effectiveness of GateSleepNet, achieving high performance in classifying sleep stages and outperforming existing methods. The proposed framework provides a powerful solution for clinical and research applications in sleep medicine, with potential for broader adoption in automated health diagnostics.