A sleep stage classification method using deep learning by extracting the characteristics of frequency domain from a single EEG channel

In the case of automatic sleep stage classification, most existing methods rely on manual functions selected from polysomnography records. This paper aims to eliminate the need for manual feature extraction by developing a deep learning-based method using single-channel EEG that automatically utilizes the time-frequency spectrum of EEG signals. The time-frequency image of the EEG signal is extracted using continuous wavelet transformation (CWT). Learning from GoogLeNet, an optimized convolutional neural network, is used to classify these CWT images into sleep stages. The proposed method is evaluated using the Haaglanden Medical Center dataset using single-channel EEG C3, C4, F4, and O2 channels. Wavelet GoogLeNet using raw data on a single EEG F4 channel achieved the highest accuracy of 77.6%. Evaluation results show that EEG signals in the front area are more useful for sleep stage classification than EEG signals in the parietal and occipital areas. This study shows the possibility of sleep stage classification of a single EEG signal without preprocessing.