MMFNet: A multi-branch multi-scale framework with adaptive sparse self-attention and cross-modal fusion for sleep stage assessment

Accurate sleep stage classification serves as a crucial foundation for sleep health assessment and disease diagnosis. However, existing approaches still encounter several challenges, including limited feature representation, redundancy in extracted information, and difficulties in effectively integrating cross-modal physiological signals. To address these issues, we propose a novel framework entitled Multi-Branch Multi-Scale Fusion with Adaptive Sparse Self-Attention and Cross-Modal Integration (MFFNet). Specifically, considering the prominent time-frequency characteristics of sleep signals, multiple branches are designed to capture multi-scale information, including a time-granular feature extraction branch and a time-frequency extraction branch. Furthermore, a multi-head adaptive sparse self-attention mechanism is introduced to suppress redundant information while emphasizing discriminative features. In addition, we employ an adaptive cross-modal fusion strategy that dynamically integrates information from EEG and EOG, and further visualize the contribution of each modality to sleep stage classification. Experiments conducted on the Sleep-EDF-39 and Sleep-EDF-153 datasets demonstrate the effectiveness of the proposed approach. Using the Fpz-Cz EEG channel and EOG signals, MFFNet achieves accuracy rates of 84.26% and 81.86%, with F1 scores of 75.91% and 73.38%, respectively, highlighting its competitive performance in sleep stage assessment.