Research On Automatic Sleep Staging Algorithm Based On Self-attention

Sleep,as an essential physiological activity of the human body,has attracted much attention.Sleep disorders in modern society seriously affect people’s health and life.The sleep stage was the foundation of the sleep related research,and it was an important evidence for the study of sleep characteristics and the diagnosis of brain diseases.Manual staging was a time-consuming and subjective task,and traditional machine learning is a complex and difficulty task.With the development of artificial intelligence,in the aspect of sleeping medical technology,the use of deep learning technology was more and more clinical value.Firstly,it summarizes the sleep related technologies,and expounds in detail the concept and characteristics of sleep signals(such as EEG and EEG)in this thesis.Meantime,the standard of sleep stage and the characteristics of each sleep cycle are introduced.Then it summarizes the sleep staging algorithm,including preprocessing algorithm,feature engineering algorithm and staging algorithm.Finally,the related technologies of deep learning are summarized and summarized.Then this thesis uses the method of deep learning to study from two angles.1.In this thesis,the common convolutional neural network is used to replace the feature extraction and screening in machine learning algorithm.To ensure the quantity of data,multiple sleep signals were selected to make classification data.At the same time,these signals are preprocessed using a wavelet threshold denoising algorithm,and then the sleep signal and the combined signal with other signals are sent to the established deep onedimensional convolution network model for classification,and different classification standards are proposed to verify the generalization of the model.The results show that the accuracy was higher at multiple signal inputs than at a single signal.Under multiple signal inputs,the average accuracy of two classification is 98.20%,three classification is 91.65%,four classification is 89.91%,five classification is 87.80%,six classification is 86.29%.2.Considering that the essence of sleep staging is sequential multi classification,the common convolutional neural network is only a simple multi classification task.In order to explore the influence of different channel data on sleep staging and learn the timing characteristics of sleep data,an automatic sleep staging based on self attention mechanism is proposed on the basis of the convolutional network structure in this thesis.To ensure the quantity of data,multiple sleep signals were selected to make classification data.After preprocessing,they are sent to the model based on self attention mechanism for classification.At the same time,different classification standards are used to verify the generalization of the model.The results show that under multiple signal inputs,the average accuracy of two classification is 98.87%,three classification is 95.03%,four classification is 93.58%,five classification is 90.59%,six classification is 89.15%.These results demonstrate the comprehensive performance and stability of the proposed algorithm over others.