Intelligent Analysis And Research Based On Multi-modal Sleep Signals

Sleep is an important part of body's recovery.The research of sleep stages and sleep arousal is the basis for diagnosing sleep disorders.Polysomnography is usually used to collect relevant physiological data of human body.The doctor manually labeled the polysomnography and analyzed the sleep stages and arousal.Due to the large amount of data and different data formats,the workload of this method is heavy.Due to the influenced by the doctor's personal experience,the diagnosis result is not consistent easily.With the development of artificial intelligence and big data technology,the intelligent analysis and recognition of sleep signals can be realized.This can assist the doctor's diagnosis and reduce the doctor's workload.At the same time,it can combine big data with doctor's experience,which has important clinical significance.In this paper,aiming at the detection of sleep arousal events and sleep staging in the field of automatic recognition of sleep signals.Firstly,a sleep arousal classification model Convolutional-Residual network with Positional Embedding and Multi-head Attention(CRPEMA)was designed based on multi-modal sleep signals.Based on the CRPEMA model,a multi-task deep learning model for sleep arousal and sleep staging recognition was constructed.Finally,a remote sleep analysis system is designed and implemented.The work done in this dissertation is following:1)A sleep arousal classification model CRPEMA was constructed.Firstly,the data is preprocessed by Fast Fourier Transform and global normalization.The data imbalance is solved by data generator.Then a CRPEMA integrated model is proposed,which uses multi-head attention(MA)with position embedding(PE)instead of Long Short-Term Memory(LSTM)structure to extract time series features in parallel.By improving the residual block structure,the model retains local and global features while reducing dimension.Results showed that CRPEMA obtained 0.391 Area Under the Precision-Recall curve(AUPRC)and 0.844 Area Under the Receiver Operating Characteristic curve(AUROC).2)A multi-modal multi-task sleep analysis model was constructed.Firstly,the dataset was preprocessed and the multi-task label was obtained.Then,the multi-modal multi-task model was constructed with CRPEMA.Focal loss(FL)and Cross Entropy(CE)were used as loss functions to detect sleep arousal and sleep staging to deal with multi-classification tasks under data imbalance.Finally,the effects of loss function combination,unbalanced class label weight and joint weight between loss functions are explored.The experiment showed that the multi-modal multitask sleep analysis model obtained 0.6822 macro accuracy in the task of sleep arousal classification(category 13)and 0.8122 macro accuracy in the task of sleep staging classification(category 5).3)A sleep analysis system was designed and implemented.Firstly,the demand analysis of sleep analysis system is given.Secondly,the design and implementation of user management module,data acquisition module,data analysis and processing module and front-end display module are completed.Finally,Hadoop Distributed File System(HDFS)storage environment for a large number of sleep signals is designed and implemented.The test results show that the system can assist doctors to analyze and diagnose sleep signals remotely.