Study Of Automatic Sleep Staging Based On Single-channel EEG

Sleep directly affects human health and quality of life,and accurate sleep staging is crucial for evaluating sleep quality.Automatic sleep staging can assist sleep experts in assessing sleep quality,improve sleep staging efficiency,reduce human errors,and reduce evaluation costs,contributing to the implementation of home sleep monitoring.Compared with the sleep staging using multi-channel signals,the method using single-channel EEG has the advantages of small device size,convenient wearing,and low sleep interference.Nowadays,domestic and foreign scholars use machine learning and deep learning to study the sleep stages of single-channel EEG.However,these methods still have problems such as low accuracy of feature extraction and classification in a single analysis domain,insufficient temporal context capture,and the need to improve feature learning ability.This dissertation focuses on the above difficulties and studies an automatic sleep staging method based on single-channel EEG.The aim is to provide a new method for wearable or portable devices,provide practical technology for patients to achieve long-term home sleep monitoring,and promote the development and application of automatic sleep staging in healthcare.The main contributions of this paper are as follows:1.In response to the single-channel EEG contain simple information and traditional single analysis domain feature parameter extraction algorithms cannot meet the accuracy of sleep stage classification,this dissertation adopts a sleep staging method that combines time-domain and frequency-domain features.Through empirical mode decomposition,local features of EEG signals at different time scales are obtained,and the proposed frequency-domain features are used to distinguish the characteristics of EEG signals in different sleep stage frequency ranges.The experimental results show that combining the time-frequency domain features of EEG signals can effectively improve the classification performance of sleep stages,with an accuracy rate of 74.4% in the St.Vincent’s University Hospital/University College Dublin Sleep Apnea Database.2.This dissertation analyzes the time of the target period’s EEG signal and its pre order time EEG signal,extracts and utilizes the temporal context information of the long and short periods of sleep,and enhances the model to learn the characteristics of EEG signals in each sleep stage through the data augmentation algorithm.This method not only achieved better performance in the dataset of healthy individuals,but also achieved an accuracy of 78.8% in the CAP Sleep Database composed of patients with sleep disorders.The experimental results indicate that using long and short-term sleep stage sequence temporal context can effectively improve the accuracy of sleep staging.3.Due to the abstract process of extracting features from each sleep stage using convolutional neural networks in traditional methods,which may result in the loss of boundary context,a sleep staging method based on boundary temporal context enhancement is proposed.This dissertation refines the boundary information of sleep stages based on extracting multi-scale time dependencies,enhancing the model’s learning ability for boundary temporal context during sleep stage transitions.At the same time,a sequence-based data augmentation method with category awareness is designed to improve the model’s ability to learn the boundary temporal context between a few categories of sleep stages and other sleep stages.The experimental results show that using boundary enhanced multi-scale temporal context can learn sleep stage transition relationships,achieving an accuracy of 85.2% in the Sleep Heart Health Study.4.Considering the feature reconstruction of EEG during a sleep stage,ignoring the role of sequence temporal context in capturing long-term dependencies and enhancing representation,this dissertation proposes a multi-task sleep staging method that combines sequence signal reconstruction and sequence signal segmentation.This method can optimize the common encoder through sequence signal reconstruction,enhance the representation learning of EEG signals including sequence temporal context,and achieve an accuracy of 85.6% on the Sleep-EDF Database Expanded 2013 dataset.The experimental results show that the generalization ability of the model can be improved and the risk of overfitting can be reduced by using sequence signal reconstruction.