Research On Sleep Stage Classification Combined With Feature Learning And Sequence Learning

Sleep is an important part of human life.The sleep quality is closely related to peoples physical and mental health.Sleep stage classification is a technique for analyzing sleep quality.And it is also an important reference for the diagnosis and treatment of sleep disorders.Manual sleep staging is the golden standard of sleep stage,but it needs sleep experts to spend a lot of time processing sleep data.So there is an urgent need for a reliable automatic sleep staging method,which can not only reduce the work of sleep experts,but also promote the development of portable sleep monitoring equipment.As the sleep stage transition rule is an important reference for sleep physicians to label,this thesis proposes two automatic sleep staging algorithms which combine feature learning and sequence analysis based on machine learning and deep learning methods,focusing on how to exploit the transition rule to improve the performance of the algorithm.The research of this thesis mainly includes the following aspects:(1)GBDT-CRF model: Firstly,according to the characteristics of signals in different sleep stages,time-domain,time-frequency domain and nonlinear features are extracted from EEG,EOG and EMG channels to form the feature vector of each epoch.Secondly,gradient boosting decision tree(GBDT)is trained to fit the data and pre-predict the sleep stage.Thirdly,as conditional random field(CRF)is good at learning the transition rule of hidden state chain,it is trained to modify the pre-prediction result and output a more reasonable staging sequence.(2)Dense Net-LSTM model: Firstly,construct a Densely Connected Convolutional Networks(Dense Net)to learn the features automatically.This network reuses the features extracted from the front layer through dense connection.With the superposition of the convolution layer,it can effectively fuse the time-frequency features and compress signal as low-dimensional feature vector.Secondly,the Long Short-Term Memory(LSTM),which is good at dealing with the sequence dependence,is used to analyze the transition relationship.In order to improve the networks attention to the long-term dependence,attention mechanism is added in the LSTM encoding process.Finally,experiments on two public data sets(Sleep-EDFx and Physionet2018)show that the method proposed in this paper can achieve stable and reliable automatic sleep staging.Also,the data analysis results show that the feature vectors described in this thesis are effective.Adding sequence analysis to the model can significantly improve the accuracy of sleep staging.Combining feature learning and sequence analysis is a practical method for sleep staging.