A Two-Phase Model for Sleep Staging Using Single Channel EEG

Abstract

Sleep staging is the basis for assessing sleep quality and diagnosing sleep disorder. In this paper, we propose a two phase model, based on deep neural networks and support vector machine for automatic sleep staging using raw single channel EEG signals. Instead of hand-engineering features, the first phase of model apply the combination of convolutional neural networks (extract time-invariant features) and bidirectional long short-term memory networks (learn temporal correlation among sleep stages) to learn features automatically from raw EEG. The second phase of model use traditional support vector machine classifier to identify 5 sleep stages based on previous extracted features in the first phase. Compared to existing methods, our model can learn richer features from raw EEG signals automatically, and achieves a better accuracy of 78.34%. The model is evaluated on the Sleep-EDF database, and we use independent training and test datasets. The experiments show our model achieves an excellent result.