Research On Sleep Apnea Detection Method Based On ECG

Sleep apnea syndrome is a kind of sleep disease with wide influence and can cause multiple organ diseases.Since many studies have found that it has a high correlation with ECG signals,the use of ECG signals to detect sleep apnea has become a research hotspot.However,the existing ECG-based detection algorithms still face many problems,including the need for complex feature engineering,the lack of consideration of ECG temporal features of sleep apnea,and the inability to adaptively preprocess the ECG signal.In view of the above problems,this thesis starts from the direction of deep learning,and the main research contents are as follows:（1）In view of the problem that the existing detection algorithms need complex feature engineering and rely heavily on empirical parameters,this thesis proposes a sleep apnea detection model based on a two-stream architecture.This thesis uses the ECG signal as the first input,and uses the ability of convolutional neural network to abstract local details and the ability of bidirectional GRU to extract contextual timing information for feature extraction.Since the R-R interval sequence was proven to be useful for sleep apnea detection,it was used as a second input to automatically extract features associated with sleep apnea syndrome discrimination.Through the experiments on the Apnea ECG public dataset,it is verified that the proposed algorithm achieves a detection level similar to the existing algorithm,and the importance of the two feature extraction modules is verified by ablation experiments.（2）Aiming at the problem that the existing detection algorithms do not consider the ECG temporal features of sleep apnea,this thesis proposes a sleep apnea detection network based on the temporal attention mechanism.For the first time,an attention mechanism is introduced in this research field to capture time segments related to sleep apnea.By assigning more attention weights to relevant time segments,the information characteristics of relevant intervals are enhanced,and the interference of irrelevant intervals is reduced.Through the visualization of the intermediate results of the model and the comparison with the performance indicators of the existing algorithms,it is verified that the proposed model can effectively capture the time interval of sleep apnea,and the detection effect reaches a level comparable to the existing algorithms.（3）Aiming at the problem that the existing algorithms cannot adapt to various noise interference,this paper explores the role of the adaptive preprocessing network based on Unet++ in the denoising of ECG signals,and then proposes an end-to-end sleep apnea detection with the addition of an adaptive denoising subnet method.The ECG signal is usually interfered by a variety of noises.Although the method based on artificial experience parameters plays a certain role in denoising,its parameters are easily affected by noise.Unet++ can capture and fuse the fine-grained and coarse-grained temporal information of the signal,which is more conducive to the adaptive denoising of ECG.Through the experiments on the database constructed based on the Apnea ECG dataset and the MIT-BIH noise stress test dataset,it is verified that the adaptive preprocessing network proposed in this paper has good denoising ability.At the same time,through comparative experiments,the end-to-end network adding the adaptive preprocessing subnet reduces the balanced error rate by 1.36%on average compared with the existing algorithm,and can achieve better detection results.