Research On Sleep Stage Classification Based On EEG Time-Frequency Information Using Deep Learning Networks

Sleep disorders are a type of condition that disrupts the normal patterns of sleep and wakefulness,leading to poor sleep quality.Common sleep disorders include insomnia,hypersomnia,sleep apnea,and circadian rhythm disorders.Sleep staging is an effective method for diagnosing sleep disorders.Traditional sleep staging methods primarily rely on the analysis of data such as electroencephalogram(EEG),electrooculogram(EOG),and electromyogram(EMG)signals by expert physicians.However,analyzing hours-long sleep data can be highly labor-intensive for physicians,and the quality of sleep staging is also greatly influenced by their expertise.To address these issues,this paper proposes three automatic sleep staging methods based on deep learning.Firstly,this study utilized data from 20 volunteers in a publicly available dataset,with their frontal electroencephalogram(EEG)signals as the analysis target,to stage their sleep states.To alleviate the impact of data imbalance on the model,a balanced dataset of sample categories was constructed using oversampling.Initially,23 common features from the time domain,frequency domain,and time-frequency domain were extracted from the EEG signals,and traditional machine learning methods were employed for sleep staging.However,traditional machine learning approaches struggle to effectively handle the temporal information within EEG features.Therefore,this paper proposes a deep learning method based on EEG features,consisting of residual networks and bidirectional Long Short-Term Memory(LSTM)networks.This method can consider both past and future contextual information when processing EEG feature data,enabling a more comprehensive understanding of the input feature sequence.Compared to traditional machine learning methods,this approach demonstrates higher performance in sleep staging,achieving an accuracy of 85.8%.Furthermore,the process of EEG feature extraction heavily relies on manually designed extractors,requiring domain expertise and complex parameter tuning.Additionally,each method is specific to a particular application,resulting in poor generalization and robustness.Therefore,this paper proposes an end-to-end sleep staging model based on deep learning,called Time-Frequency domain Self-Attention Temporal Network(TFSTSleep Net).This network employs multiple branches to process EEG signals,as well as the frequency and time-frequency signals obtained from processing the EEG signals,to extract more comprehensive signal features through convolution.The combination of bidirectional LSTM and self-attention helps capture the dependencies between sleep stages before and after a given time point more effectively.To validate the effectiveness of the three branches,bidirectional LSTM,and self-attention module,a comparison of sleep staging was conducted between a single-branch residual network model,a three-branch time-frequency domain residual network model,a three-branch time-frequency domain bidirectional LSTM network model,and the TFSTSleep Net model.The results demonstrate that the proposed TFSTSleep Net model achieves the best classification performance with a staging accuracy of 87.2%.Furthermore,to verify the effectiveness of each branch,ablative experiments were performed,indicating that the short-time Fourier transform branch and the EEG signal branch play significant roles in sleep staging.Finally,to further extract prominent waveforms and sequential information from EEG signals in sleep stages,this chapter proposes a Multi-Task Sequence Sleep network(MTSSleep Net).The model consists of a Time-Frequency domain Multihead Attention module(TFM),a Stage Estimation Block(SEB),and a Context Encoder Block(CEB).The results demonstrate that this model exhibits superior performance in sleep analysis,achieving an accuracy of 86.7%.This performance surpasses some existing models such as Trans Sleep Net(86.5%),Seq Sleep Net(85.2%),and Res Net MHA(84.3%).In summary,this paper presents three different sleep staging architectures utilizing single-channel frontal EEG signals.These methods demonstrate high accuracy,with the MTSSleep Net model surpassing some existing sleep staging approaches.The proposed methods contribute to the automation of sleep staging,effectively reducing the workload of physicians.