MFSleepNet: A multi-receptive field sleep networks for sleep stage classification

Sleep stage classification is essential for assessing sleep quality and diagnosing sleep disorders. However, most existing deep learning-based methods extract features from each channel's electroencephalogram signals, which overlook the spatio-temporal features of different channels. Therefore, making full use of the spatiotemporal features is still a challenge. To tackle this challenge, we propose a multi-receptive field sleep network (MFSleepNet) to capture different levels of graph structure features. This network includes the feature extraction module, an enhanced spatio-temporal feature module, a multi-receptive graph convolution network, and an attention fusion module. The feature extraction module obtains rich features through feature augmentation based on features at different frequencies. An enhanced spatio-temporal feature module is designed, which mainly includes a temporal gating layer, temporal attention, and spatial attention. This module can extract useful temporal and spatial features. In addition, the multi-receptive graph convolution network module is used to extract structural features at different levels. Then, we use the attention fusion module to learn global information to selectively emphasize informative features and suppress less reliable features. We validate the effectiveness of the proposed framework on the ISRUC-S3 dataset. The overall performance is better than the baseline method. This method can potentially bean effective tool for quickly diagnosing sleep disorders.