MHA-SleepNet: A Multi-scale Hybrid Attention Network for Sleep Stage Classification Based on Single-Channel EEG

Sleep staging has attracted much attention as an important method for studying sleep disorders in recent years. The majority of the current automatic sleep staging methods focus on studying time-domain information and ignore the interrelation between features, resulting in low sleep classification accuracy. To solve these problems, a multi-scale hybrid attention network named MHA-SleepNet is proposed for automatic sleep stage classification, using single-channel electroencephalogram (EEG) signals. The network consists of a multi-scale feature extraction (MFE), residual squeeze and excitation network (RSE), and a multi-head gated perceptron (MGP). The MFE module uses convolutional kernels of different sizes to fully extract different scale features from EEG signals. The RSE module further optimizes the weight of features and improves the feature expression ability of the network. The MGP module uses multi-head attention mechanism to capture temporal dependencies between features. The MHA-SleepNet model is evaluated on two public datasets, Sleep-EDF-20 and Sleep-EDF-78. It achieves the accuracy of 86.1% and 83.2% on the Fpz-Cz channel, respectively. Compared with the existing sleep staging methods, our method improves the classification performance.