Research On Automatic Sleep Stage Classification Method Combined With Time-frequency Information And Deep Learning

Human sleep can usually be divided into five stages,namely wake stage(Wake),the first stage of non-rapid eye movement(N1),the second stage of non-rapid eye movement(N2),slow wave sleep(SWS),and rapid eye movement stage(REM).Sleep stage classification is one of the important means to study sleep,and it is of great significance for the diagnosis and treatment of sleep diseases.At present,sleep stage classification is mainly dependent on the manual interpretation of professional physicians.The process of manual interpretation is not only time-consuming and subjective.Furthermore,the staging accuracy of manual interpretation is not high.The automatic stage classification algorithm can effectively reduce the staging error and improve the efficiency of classification.Therefore,the related research of automatic stage classification has important clinical application value.In recent years,deep learning models(such as Convolutional Neural Network(CNN)and Long Short-Term Memory Network(LSTM))have been developing rapidly.These models can not only automatically learn the effective feature representation of data from a large number of input data,but also reduce the errors caused by artificial features.Therefore,the main research goal of this paper is to build an automatic stage classification model based on the time-frequency information of single-channel electroencephalogram(EEG)and the learning ability of deep learning model.The automatic stage classification models constructed in this paper are to automatically identify the sleep stage of single-channel EEG signal every 30 seconds.Due to the strong nonlinear and nonstationary characteristics of EEG signals,how to effectively analyze and transform it to make it suitable for models of CNN and LSTM is a challenge in this study.This paper has carried out the research from the following aspects.First of all,based on the one-dimensional temporal features of EEG,the automatic stage classification model(IDT-LSTM)is constructed by using the advantages of LSTM in dealing with long-distance dependence in time series data.Secondly,a method of transforming one-dimensional EEG signal into two-dimensional temporal feature matrix is proposed in this paper,and then we constructed an automatic stage classification model 2DT-CNN based on the advantage of CNN model for pattern recognition of two-dimensional data.Finally,three time-frequency analysis techniques of signal,such as Hilbert-Huang Transform(HHT),Wavelet Transform(WT)and Short-time Fourier Transform(STFT),are used to transform the one-dimensional EEG signals into two-dimensional matrix of time-frequency energy spectrum,and we further built three automatic stage classification models with the CNN model:HHT-CNN,WT-CNN,and STFT-CNN.The time spectrum energy matrix obtained by HHT can effectively describe the local subtle changes in time and frequency of EEG signals at different sleep stages.Because the model of HHT-CNN combines the advantage of HHT’s time-frequency analysis and the ability of feature learning of CNN,it has great research and application value.Based on the 39 overnight sleep EEG data from PhysioNet,the national physiological database of the United States,we used five automatic stage classification models to carry out the studies of automatic stage classification.The experimental results show that most of the models have achieved the classification accuracy rate of more than 80%,and the model HHT-CNN achieves the best classification accuracy rate of 84.5%.Different from the traditional automatic sleep stage classification methods based on artificial feature construction,the models proposed in this paper are simpler and more efficient without designing various features of the input of original EEG signal.Meanwhile,the design ideas of the models are suitable for analyzing nonlinear and nonstationary time series,and also provide a new research idea for other classification problems based on nonlinear and nonstationary signals.