Automatic Sleep Staging Based On Deep Learning Classification Method

Sleep staging is the crucial step for physicians to evaluate the sleep quality of patients and diagnose sleep disorders and other related diseases.At present,it mainly relies on physicians manually reviewing and interpreting sleep recordings.However,manual interpretation has limitations such as time-consuming and high requirements for physician experience.Meanwhile,due to the limited level of human knowledge of how the brain works during sleep,the results of manual interpretation are subjective.Therefore,it is of great research value to study efficient,consistent,and objective automatic sleep staging methods.This paper focuses on the deep learning classification method of automatic sleep staging.The main work are as follows:(1)To address the issue of how to combine the characteristics of the sleep period with the model design in the many-to-many mode,an automatic sleep staging model based on the window attention mechanism is proposed.The overall structure consists of four modules:a feature extraction module based on a multi-branch convolutional neural network,a feature refinement module based on channel attention,a temporal capture module based on the window attention mechanism,and a linear classifier.To make use of the physician’s experience in interpreting sleep stages,making the model pays more attention to the feature channels within the limited range of the window,a temporal capture module based on the window attention mechanism is proposed.Meanwhile,to handle the increment of output feature channels in the many-to-many mode,downsampling and channel attention are used to retain key feature channels and discard other lowimportance feature channels.According to the experimental results on public datasets,the method achieves a better sleep staging outcome.Furthermore,through qualitative analysis of the attention matrix,it is demonstrated that the proposed method widens the gap between the inside and outside of the window in attention scores.(2)Aiming at how to use unlabeled data to construct a robust feature representation for the input signal,this research proposes an unsupervised automatic sleep staging algorithm based on mask learning.The model in the pre-training phase consists of three parts:a feature extraction network implemented by a single convolution layer,a Transformer-based encoder,and a decoding network composed of a bottleneck layer and a decoder.To improve the predictive power of the decoding network,a bottleneck layer is employed to decouple the encoder width from the decoder width.At the same time,considering the nonlinear and unsteady characteristics of physiological signals,it is proposed to use the features extracted by the feature extraction network in the iterative process as the prediction target,instead of using the original input as the prediction target in previous studies.According to the experimental results of the model under the linear probing evaluation mode,the method outperforms the algorithm that uses the raw signal as the target for prediction.Moreover,the method outperforms supervised baseline models on large-scale datasets and even can be further improved by widening the decoder with bottleneck layers.