Research On Sleep Apnea Detection Model Based On ECG Signal

Good sleep is an important guarantee for a normal life.Among various sleep diseases,sleep apnea hypopnea syndrome(SAHS)is a disease with a very high incidence,which has a negative impact on people’s short-term and long-term health.At present,the detection process of the disease is cumbersome and expensive,and requires professional operation by medical staff,so many people cannot timely understand their sleep status.If the detection process of sleep apnea is simplified,more people can understand their sleep status,which is very important for the prevention of diseases.In recent decades,a large number of studies have found that electrocardiogram(ECG)signals can detect the occurrence of sleep apnea,but there are still many problems to be studied in depth.This paper is to establish a diagnosis model of sleep apnea on the basis of studying the stable feature selection,to simplify the diagnosis of sleep apnea.This paper will implement the detection of sleep apnea events based on the ECG signal.The main work includes the following aspects:(1)Feature extraction and significant difference analysis.The electrocardiogram-derived respiration(EDR)signal is derived by the area of the QRS wave.The RR interval sequence and EDR signal were analyzed in time domain,frequency domain and nonlinear,and 46 features were obtained.The Mann-Whitney U nonparametric test was used to analyze the significant differences between normal sleep signals and sleep apnea signals,and the features with no significant differences were deleted.(2)Study on stable feature selection.Based on three good feature selection methods:minimum redundancy maximum correlation(mRMR)method,ReliefF method,ILFS method and a robust rank aggregation(RRA)method,four stable feature selection methods are established:ReliefF-RRA method,mRMR-RRA method,ILFS-RRA method and FP-RRA method,of which FP-RRA method is to integrate the above three basic sorting methods together.The results show that the ReliefF-RRA method and the mRMR-RRA method are more stable than the ILFS-RRA method and the FP-RRA method.Among the top 20 features of the four methods,the common features are all from RR interval,include:RRcorr2,RRcorr3,RRcorr4,and FMEn.These features have good stability and distinguishing ability.(3)Research on the diagnosis model of sleep apnea.The ranking results of the above four methods are applied to RBF-support vector machine(RBF-SVM),k-nearest neighbor(KNN)and linear discriminant analysis(LDA)classifiers to observe the change of classification accuracy with the number of features.Combine the best performing feature selection method with the classifier to form the final classification model,and apply it to the independent test set for performance verification.The results show that the RBF-SVM performs best among the three classifiers.The performance of ReliefF-RRA method and FP-RRA method is better than mRMR-RRA method and ILFS-RRA method.Combining the stability of the feature selection method and its performance on each classifier,this paper combines the ReliefF-RRA method with RBF-SVM to train the final classification model.The first 33 features obtained by the ReliefF-RRA method achieved an accuracy(Acc)of 92.1 7%,a sensitivity(Se)of 87.70%,and a specificity(Sp)of 94.46%on the test set.