Research On EEG Signal Representation Learning For Sleep Staging

Sleep staging based on electroencephalogram(EEG)plays an important role in the clinical diagnosis and treatment of sleep disorders.In order to emancipate human experts from heavy labeling work,deep neural networks have been employed to formulate automated sleep staging systems recently.However,supervised learning has been widely used in the field of EEG signal analysis,EEG signals lose considerable detailed information in network propagation,which affects the representation of deep features,and representation performance of neural networks is limited by the amount of annotated data.To address the problem of information loss,a residual block-based network and a multi-scale feature fusion model are combined to achieve feature extraction.To reduce the output jitters generated by the classifier,the Markov-based sequential correction algorithm is designed.To make up for the data shortage of supervised learning,selfsupervised learning is introduced to obtain a pre-trained model from a large amount of unlabeled data.Through the above design and optimization,the proposed method enables the model to have better EEG signal representation ability,and thus can outperform the existing methods in sleep staging tasks with multiple open datasets.The main work and contribution points of this paper are summarized as follows:(1)A new framework is proposed for data-driven sleep staging.The backbone network is a residual block-based network,which performs as a feature extractor.Then the fusion model constructs a feature pyramid by concatenating the outputs from the different depths of the backbone,which can help the network better comprehend the signals in different scale.Feature fusion is bound to be accompanied by the repetition of information.In order to reduce the redundancy of information,an adaptive channel integration module is added into multi-scale feature fusion model.The attention of different channels is calculated for weighting the original inputs to weaken common characteristics and enhance the unique characteristics,so as to improve the effectiveness of feature fusion.By retrieving the lost information of the signal,the representation ability of the network is improved,and the accuracy of sleep staging in Sleep-edfx dataset is 83.42%.(2)The Markov-based sequential correction algorithm is designed for the classifier post-processing.CNNs lack of the ability of capturing temporal informantion and regard input signals as independent identically distributed,so as to predict sleep stages may change many times in a short duration,namely,’jitters’.In order to solve this problem,we design a sequence correction algorithm based on Markov chain based on the physiological law of sleep cycle and the correlation between the current stage,the previous stages and the incoming stages.The algorithm depends on a prior stage distribution associated with the sleep stage transition rule and the Markov chain.By using this post-processing algorithm,the problem that the model could not obtain time sequence information is solved,and the accuracy of sleep staging in Sleep-edfx dataset is improved by 1.72%.(3)A self-supervised contrastive learning method of EEG signals is proposed for sleep stage classification in this paper.EEG signals are usually simple to obtain but expensive to label.Self-supervised learning is introduced to reduce the labeling cost of EEG signals and improve the representation capability of the network.During the training process,a pretext task is set up for the network in order to match the right transformation pairs generated from EEG signals.In this way,the network improves the representation ability by learning the general features of EEG signals.The robustness of the network also gets improved in dealing with diverse data,that is,extracting constant features from changing data.In detail,the network’s performance depends on the choice of transformations and the amount of unlabeled data used in the training process of self-supervised learning.Compared with the randomly initialized model,the pre-trained model obtained through self-supervised learning improved the accuracy of sleep staging by about 3%.