<p>Sleep stages refer to the distinct processes within a person’s sleep cycle. They are essential for assessing mental and physical health. Existing sleep stage classification models typically improve performance through increased computation complexity or be trained with more labeled data. These models may result in overly heavy models that are unrealistically not applicable in real-world scenarios. To address this issue, this paper proposes a Spatial-Channel Attention Network for Sleep Stage Classification&#xa0;(SCANSleepNet). This framework is built on the time-frequency characteristics of EEG signals and the conversion rules of sleep stages. It constructs a lightweight deep-learning system that integrates multi-scale frequency analysis and dynamic feature enhancement. Its improvements lie in two main aspects. First, a Spatial-Channel Dimensional Attention&#xa0;(SCDA) block is designed to model the dynamic transition among stages while requiring fewer parameters. Second, a weighted cross-entropy loss function is introduced to address class imbalance without dependence on extra data augmentation. It enhances the model’s lightness and suitability for clinical applications. Experimental results show that the SCANSleepNet achieved 85.52% accuracy in the Fpz-Cz channel and 82.16% in the Pz-Oz channel on the Sleep-EDF dataset. It makes a good balance between classification accuracy and efficiency.</p>