Abstract Background This study presents SIMSleepSM, a novel single‐channel electroencephalography (EEG) sleep staging model. It addresses two primary challenges: insufficient modeling of long‐range temporal dependencies combined with limited multi‐scale feature extraction, and poor accuracy in identifying the N1 stage. Methods SIMSleepSM extends the SleePyCo architecture through three principal innovations. First, the spatial‐channel synergistic attention (SCSA) module is adapted into a 1D variant, SCSA₁D, tailored for EEG signals and inserted into every feature layer of the backbone network. The spatial attention extracts local temporal dependencies across various time scales, whereas the channel attention captures relationships among feature channels. Together these attentions strengthen temporal dependency modeling and emphasize N1‐specific features. Second, an interactive convolution block (ICB) is integrated into the feature pyramid. The ICB employs a two‐branch interactive convolution to refine multi‐scale feature extraction. Finally, a bidirectional Mamba‐based classifier is designed. Its bidirectional state space mechanism captures long‐range temporal dependencies in the EEG and thereby strengthens representation of sleep‐stage dynamics. Results On the Sleep‐EDF‐20, Sleep‐EDF‐78, and Sleep Heart Health Study (SHHS) datasets, SIMSleepSM achieves accuracy values of 88. 1%, 86. 2%, and 84. 1%; records macro F1 scores of 82. 7%, 81. 0%, and 77. 9%; obtains Cohen's Kappa coefficients of 0. 839, 0. 810, and 0. 791; and attains F1‐scores on the N1 stage of 53. 7%, 54. 4%, and 50. 9% for Sleep‐EDF‐20, Sleep‐EDF‐78, and SHHS, surpassing the second‐best models by 1. 3%, 4. 0%, and 4. 8%, respectively. Conclusion Experimental results demonstrate that SIMSleepSM outperforms thirteen state‐of‐the‐art baseline models, with particularly notable improvements in N1‐stage identification. These results indicate that SIMSleepSM provides an effective and reliable solution for automatic sleep staging using single‐channel EEG, highlighting it as a robust and high‐performing model.
Xu et al. (Fri,) studied this question.