Learning a convolutional neural network for sleep stage classification

Nan Liu, Zongqing Lu, Bokun Xu, Qingmin Liao

Abstract: This paper presents a novel automatic sleep stage classification method based on the convolutional neural network (CNN) with Electroencephalogram (EEG). To fully utilize the local frequency domain information of original EEG signals, we define the fractional discrete Fourier transform (F-DFT). For a better classification performance in the sleep stage 3 and the stage 4, wavelet transform (WT) is used to depict the low frequency structure information of local signals rather than traditional filter methods. Using the locally corresponding relation between time and frequency domains, we generate the three-dimensional signal consisting of the EEG signal, the F-DFT signal and the WT subband signal. We feed the CNN with expanded three-dimensional signals to extract features and classify. Unlike previous methods, our process of feature extracting is automatically. As far as I can see, we first introduce the CNN to sleep stage classification, and achieve state-of-the-art performance.