Adaptive Multi-Scale Convolution and Local Attention for single-channel EEG sleep staging

Abstract

Sleep staging is a critical component of sleep medicine, providing indispensable insights into sleep disorders and broader health outcomes. Single-channel EEG sleep staging offers a less invasive and more practical alternative to multi-channel methods. However, achieving high classification accuracy remains challenging, given the complex, non-stationary nature of EEG signals. In this study, we propose a new model that combines Adaptive Multi-Scale Convolution (AMSC) and Local Attention mechanisms. The proposed AMSC module captures temporal features across multiple scales, enabling the model to adapt to signal dynamics and focus on relevant frequency bands. This adaptability improves handling of the heterogeneity inherent in EEG data. The Local Attention module refines feature extraction by concentrating on the most informative segments of the signal, thereby improving the model’s capability to distinguish among various sleep stages. Our approach integrates normalization layers, a Multi-Layer Perceptron (MLP), and a Softmax classifier to ensure robust learning and accurate stage classification. Experimental validation on publicly available EEG datasets and self-collected datasets shows that the proposed model surpasses current state-of-the-art methods. Specifically, our model achieves an overall accuracy of 86.2% on the Sleep-EDF dataset, with a notable 59.1% accuracy for the challenging N1 stage. On our self-collected CQSH-OSSD dataset, the model achieves an overall accuracy of 79.4%, demonstrating strong generalization performance across different data sources. The proposed framework confirms the potential of combining adaptive convolution architectures with attention-based feature enhancement. It improves the effectiveness of sleep staging based on single-channel EEG.