Research And Application Of Multimodal Sleep Staging Algorithm Based On U~2-net

Sleep staging is the basis for the diagnosis and treatment of sleep-related diseases,and an important means of monitoring sleep status.In the traditional sleep staging method,doctors manually mark the physiological data on polysomnography and distinguish sleep stages according to sleep staging criteria,which is requires high clinical experience of the doctor,with the time-consuming and labor-intensive.Most of the existing automatic sleep staging methods use the time-domain and frequency-domain features of the original signal to perform sleep staging.This method requires certain prior knowledge,and has shortcomings such as large amount of features to be extracted and low classification accuracy.In addition,most of the existing research on sleep staging algorithms are based on single-modal EEG signals for model building,ignoring the sleep information provided by multi-modal signals in specific sleep stages.Therefore,this thesis studies the multimodal sleep staging algorithm based on characteristic waveform detection,designs and develops a polysomnography system,improves the accuracy and efficiency of automatic sleep staging,and reduces the difficulty of sleep monitoring.The main contents of the thesis are as follows:（1）Signal preprocessing.Using the method of frequency domain filtering,a low-pass filter is designed to filter out high-frequency noise such as power frequency interference and EMG in the EEG and EOG signals,retaining the effective components in both signals;This thesis use a trend term removal algorithm based on complete ensemble empirical mode decomposition with adaptive noise,making use of its adaptive filtering characteristics,to reconstruct the EEG signal and remove the low-frequency artifacts that overlap with the EEG rhythm wave in the EEG signal.（2）Research the detection method of the characteristic waveform of sleep signal.A characteristic waveform detection network based on U~2-net is designed to detect the characteristic waveforms of EEG and EOG signals in specific sleep stages,capture complete sleep signal information,and improve the accuracy of sleep staging algorithms.A multi-scale feature extraction module is adopted in the network to extract sleep transition rules at different scales to further improve the classification performance of sleep stages.（3）Research the multimodal sleep staging algorithm model based on U~2-net.The model consists of two-channel input structure with EEG and EOG signals,and adopts a multi-modal feature extraction module to fuse the waveform features of the two input signals.In this module,the dual attention network structure of channel and spatial attention is added to enhance the important waveform features of the signal in a specific sleep period,improving the classification accuracy of sleep stages.The model has been trained on two data subsets of sleep-EDF successively,and the F1 scores of W,N1,N2,N3,and REM sleep periods,respectively on the ST（Sleep Telemetry）data set,are predicted to be 93.19%,51.82%,85.96%,79.7%and 88.49%.The overall accuracy rate reached 85.06%;The F1 scores on the SC（Sleep Cassette）dataset were 92.46%,59.47%,90.2%,84.71%,91.65%,and the overall accuracy rate reached 88.12%.（4）Design a polysomnography system to realize automatic sleep staging.The system can continuously collect the subjects’nighttime sleep EEG and EOG signals,and has the functions of real-time waveform display and historical review.The system deploys the automatic sleep staging model designed in this thesis on the algorithm server,and realizes automatic sleep staging through the interaction between the software terminal and the algorithm server,which has certain practicability.