Research On Automatic Sleep Staging Based On CNN-BiGRU

Sleep staging is to divide a person’s entire sleep process into several stages according to certain standards.As an important part of sleep medicine research methods,sleep staging provides basic data for the diagnosis of sleep diseases based on a reasonable evaluation of sleep quality.For sleep staging,many scholars initially achieved good staging results according to the manual classification of sleep experts,but it was time-consuming and labor-intensive,and it had high requirements for the experience and ability of experts.In recent years,with the increasing popularity of technologies such as machine learning and deep learning,sleep staging based on deep learning methods such as convolutional neural networks has achieved good results,but the accuracy and efficiency of staging need to be further improved.Therefore,the research on automatic sleep staging based on convolutional neural network in this thesis has certain theoretical and practical significance for enriching sleep staging methods of EEG signals and promoting the progress of sleep medicine research.The research on automatic sleep staging based on CNN-Bi GRU carried out in this thesis mainly includes: sleep EEG data preprocessing,sleep feature extraction and temporal information extraction.Firstly,the raw sleep EEG data is filtered and class-balanced,secondly,sleep features are extracted by convolutional neural network（CNN）,and a bidirectional gated recurrent unit（Bi GRU）is added to learn the temporal rules between sleep stages.Softmax classifier for sleep staging.The main research work of this thesis is as follows.（1）Aiming at the problem of how to remove other interfering signals and data sample imbalance in the sleep signal obtained by the polysomnography detection system,this thesis adopts the Butterworth bandpass filter for filtering processing and the synthetic minority oversampling technique for class balance processing.The original sleep EEG signal was extracted with a band-pass filter from 0.4Hz to 30 Hz,and the data samples were adjusted by synthetic minority oversampling technique in the model training stage,which provided the basis for accurate sleep staging.（2）Aiming at the problem of how to extract multi-level features of sleep EEG signals,this thesis proposes a three-scale parallel convolutional neural network method to extract features at multiple time and frequency resolutions.The preprocessed data is input into three parallel convolutional neural networks.After a series of operations such as convolution and maximum pooling,the EEG is extracted by setting different sizes and steps in the three-scale convolutional neural network.The sleep features at different levels such as detail,structure and shape in the signal are input into the next step sequence residual learning module as a feature sequence for staging.Experiments show that using 3CNN as a comprehensive feature learning module can effectively improve the accuracy of sleep stage recognition compared with using double-layer CNN as a feature extraction module.（3）Aiming at the problem of how to improve the long short-term memory network（LSTM）in the sequence residual learning module,the parameters are too complicated and affect the training efficiency.In this thesis,a bidirectional gated recurrent unit is proposed as the temporal information extraction module.The sleep feature sequence extracted by the comprehensive feature learning module is used as the input of the bidirectional gated recurrent unit.After learning the forward and backward time information of the feature sequence,the time series rules between sleep stages are obtained,and a shortcut connection Fc is used to directly the features converted to vectors are added to the output of the bidirectional gated recurrent unit as supplementary time series information.The experimental comparison with the model using bidirectional LSTM shows that the model using bidirectional gated recurrent unit network can obtain the best sleep staging accuracy in the shortest training time.