S³D-Net: Learning Disentangled Subject-Invariant Representations for EEG Sleep Staging via Spectral-Spatial-Sequential Feature Fusion

Accurate sleep staging is essential for evaluating sleep quality and diagnosing sleep disorders. However, existing algorithms often struggle with high inter-subject variability, limiting their generalizability. Furthermore, while electroencephalogram (EEG) signals exhibit both static anatomical and dynamic functional connectivity, most methods fail to capture this dual nature. To address these challenges, we propose S³D-Net (Spectral-Spatial-Sequential Disentanglement Network), a robust framework designed to extract subject-invariant representations for EEG-based sleep staging. The core of our model is a Hypno-S³ Encoder that fuses features across the spectral, spatial, and sequential domains. It uniquely models both static and dynamic inter-channel relationships using a dual-graph mechanism and leverages a Mamba-based module to capture the continuous temporal dynamics of brain states. To overcome inter-subject variability, we introduce a multi-faceted disentanglement strategy that learns subject-invariant representations by optimizing a composite loss function, including statistically interpretable constraints such as maximum mean discrepancy and causal consistency. Extensive experiments show S³D-Net consistently outperforms state-of-the-art methods on three diverse public datasets, establishing new benchmarks for generalized sleep staging with accuracies of 79.20% (ISRUC-S1), 79.35% (ISRUC-S3), and 78.38% (HMC).