| With the development of modern society,people are increasingly paying attention to sleep disorders and other sleep health problems,and there is growing interest in how to treat and prevent sleep disorders.At present,the sleep staging method based on deep learning still has the problem of low accuracy classification.This thesis delves deeper into three automated sleep staging techniques using various deep learning models such as convolutional neural network,generative adversarial network,recurrent neural network,and graph convolutional network as the underlying architecture.The primary focus of this study is as follows.1.A semi-supervised hybrid network on sleep staging algorithm was proposed to overcome the limitation of relying on the label of a given dataset when extracting features under supervised learning.In this method,an improved convolutional encoder-decoder was used to extract shallow temporal and spatial features of EEG,a bidirectional gated recurrent unit was fused as a generator to generate sleep sample features,and the deep temporal and spatial features were obtained for sleep staging through adversarial training with a discriminator.At the same time,the weighted cross entropy loss function and Hard swish activation function were used to accelerate the convergence and performance improvement after feature fusion.Experiments show that the overall accuracy of the proposed network on Sleep-EDF,ISRUC and SHHS datasets is 86.2%,83.4% and 89.3%,respectively,which is3.4%,2.68% and 2.7% higher than that of the convolutional recurrent network,respectively.2.Aiming at the problem that the sleep staging method based on graph convolution cannot represent the different degrees of influence between adjacent channels in the graph structure,a multi-layer temporal and spatial graph attention network on sleep staging is proposed.This network utilizes the interaction between graph attention networks and bidirectional gated recurrent unit in time-frequency domain and spatial domain respectively.Moreover,multiple spatial-temporal attention balance channel weights to dynamically adjust channel features.The transition stage estimator acts on confusing sleep stages,which effectively improves the limitations of graph convolutional networks in large-scale graph data.According to the experimental results,the method proposed in this study has achieved remarkable performance,with overall accuracy rates of 82.5% and 87.3% when applied to ISRUC and SHHS datasets,respectively.3.To address the problem of the inherent and insufficient number of EEG signal channels in current sleep datasets,leading to weak model generalization ability,a method of graph generative adversarial network on sleep staging was proposed.In this work,a brain graph structure centered on EEG channels is constructed,and the graph generation model fits the true connectivity distribution on each vertex.On the contrary,the resulting fake samples are discriminated by the graph discrimination model and the results are fed back to the graph generation model learning.In order to solve the limitation on traditional Softmax,a graph Softmax function based on node-level attention is introduced and combined with the spatial-temporal attention to improve the model performance.And to distinguish these features in low variance context,weighted sleep stage prediction is introduced to distinguish each other’s distortion distribution.The experimental results show that the proposed method achieved overall accuracy rates of 83.1% and 88.6% on the ISRUC and SHHS datasets,respectively,reaching state-of-the-art performance compared to other existing sleep staging models. |
