Research On Algorithm Of Sleep Staging Based On Deep Learning

Sleep is a vital physiological activity for human health.It has great research significance for medical research and clinical practice to monitor human sleep more accurately and comprehensively.Sleep stage classification is one of the key steps to effectively diagnose and treat sleep-related diseases.With the continuous development of deep learning technology,computers are more and more commonly used to recognize and process signals,and many intelligent and automated physiological signal recognition methods have been proposed.In order to improve the efficiency of physicians by using computer-aided diagnosis and treatment,this paper carries out the research of automated sleep staging based on deep learning.The specific work is as follows:Firstly,this paper investigates and summarizes the current research status and related technologies of sleep staging in the medical and computer fields,compares and analyzes the advantages and disadvantages between sleep staging algorithms based on traditional methods and sleep staging algorithms based on deep learning,and at the same time,clarifies the problems of the existing methods and the direction of development.In addition,this article also introduces the rules of sleep staging in detail,as well as the related knowledge and analysis methods of multi-modal sleep signals.Similarly,some technologies of deep learning neural networks are elaborated.Secondly,this paper proposes a sleep staging algorithm based on residual learning and multi-granularity feature fusion.Some existing sleep staging algorithms based on deep learning usually adopt the idea of local learning,which cannot mine the correlation between global information and local information.To solve this problem,this algorithm proposes two key technologies are named the multi-granularity feature residual learning module and the attention perception fusion module respectively.Among them,the multi-granularity feature residual learning module is a multi-branch deep network which uses three branch networks do not share net weights to extract the multigranularity features of the input data.It can help improve the granularity diversity of the features.Each branch of the module is built with the residual block as the basic unit,so the network performance will not decrease as the number of layers deepens.Furthermore,the attention perception fusion module is used to assign network perception weights to features with different granularities.Experiments prove that the proposed module improves the classification performance of the network architecture effectively.Finally,this paper proposes a sleep staging algorithm based on data adaptation and multi-modal feature fusion.The HHT algorithm is used in this method to decompose the original signal adaptively and transform it into time-frequency features to reduce the influence of the non-stationarity and individual variability of physiological signals.In order to make full use of the synergy and complementarity between different modal data,the proposed method uses two heterogeneous feature learning sub-networks to learn the advanced representation information of EEG data and EOG data respectively,and uses the multi-modal joint representation sub-network to fit the nonlinear relationship between the above-mentioned multimodal heterogeneous characteristics,so as to achieve the deep fusion between multimodal features.Experimental results show that the proposed multimodal fusion algorithm in this paper can improve the classification accuracy of single modality effectively and maintain good performance on both data sets,which proves that the model is robust.