A Deep Learning Method with Self-Attention Mechanism for Cross-Subject Sleep Stage Classification Based on EEG and EOG

Abstract

Brain-computer interface (BCI) systems based on electroencephalography (EEG) and electrooculogram (EOG) were shown to be able to be used for automatic sleep stage classification since these two signals contain many sleep characteristics. However, EEG signal characteristics vary greatly among individuals, and manual classification is time-consuming and subjective. This paper proposes a deep learning method using a self-attention mechanism to achieve cross-subject sleep stage classification based on EEG and EOG. The method mainly consists of three parts. First, a traditional convolutional neural network is used to perform preliminary feature extraction on the information of the two channels. Then use Long Short-Term Memory (LSTM) to find the features in time series. Finally, the self-attention mechanism is used to find more mission-critical information from the high-dimensional feature information. We performed 25-fold cross-validation experiments and showed that the model achieved an average accuracy of 82.4% and 80.4 of the macro-averaging F1 Score(MF1).