Heart Sound Classification Based On Order Statistics

According to the "China Cardiovascular Disease Report" issued by the National Center for Cardiovascular Diseases,the number of cardiovascular patients in China has reached 290 million.The death of cardiovascular disease accounts for more than 40% of the deaths of residents.Therefore,it is of great significance to accurately and effectively diagnose cardiovascular diseases.Current diagnostic methods include echocardiography and Magnetic Resonance Imaging(MRI).Although these techniques provide more direct and accurate evidence for the diagnosis of cardiovascular diseases,the equipment used in this type of technology It is expensive,bulky and complicated to operate.As a non-invasive diagnostic method,auscultation is a cost-effective method for cardiac examination.However,it is subjectively dependent on auscultation and requires high clinical experience for physicians.With the advent of deep learning and machine learning,ancillary diagnostic tools have been provided to reduce auscultatory dependence.Heart sound signal analysis is an effective and convenient way to diagnose the heart.However,automatic heart sound classification in signal analysis is still a challenging issue,which mainly embodies feature extraction of heart sound signals.The present proposed feature only considers the interval of the heart sound signal and the change in the time frequency,and does not consider the change of the heart sound signal amplitude.Changes in the heart sound signal amplitude reflect the heart's activity,and sequential statistics can capture this change in amplitude.Therefore,in order to extract more effective discriminating features from the classification of heart sounds,this paper proposes a heart sound classification method based on the features of sequential statistics.The main content of this study:First.Preprocessing of heart sound signals: Wavelet denoising of heart sound signals removes some noise that coincides with frequencies of heart sound signals.Heart sound segmentation: The four-component heart sound(S1),systolic phase,and first-division heart sound signal are divided into four parts based on Logistic Regression-Hidden Semi-Markov Models(LR-HSMM).Two heart sounds(S2),diastolicphase.Second.Feature extraction of heart sound signals: It includes the extraction of four features: sequential statistical features,time domain features,frequency domain features,and Convolutional Neural Network(CNN)features.Extraction of sequential statistical features(26-dimensional eigenvectors): differentially sorting the preprocessed heart sound signals to obtain a difference vector X of the heart sound amplitude,representing the numerical distribution of X using a histogram,and the statistical distribution vector(21-dimensional vectors)As a feature,take the maximum,minimum,mean,1/4 digit,and three-quarters number(5-dimensional vector)of vector X as the sampling feature for X.Time-domain feature extraction and frequency-domain feature extraction: The heart sound signal is preprocessed,then the heart sound is segmented,and the Hamming window and the discrete-time Fourier transform are used to extract the time-domain and frequency-domain features.Extraction of CNN features: After preprocessing,the cardiac cycle is extracted,band-pass filtering is used to segment the frequency band,and CNN features are extracted.Third.Classification of heart sound signals: Although Support Vector Machine(SVM)is a two-layer neural network,it can be convexly optimized to find the global optimum.Therefore,this paper selects the support vector machine as the heart sound classifier.The time domain features,frequency domain features,CNN features,and sequential statistics features are used to test and train the SVM classifier.Then compare and analyze the classification results using these feature combinations.The heart sound data used in this article is from the 2016 PASCAL Heart Sound Classification Challenge.The data set consists of five databases(A to E)and contains a total of 3,126 heart sound recordings lasting from 5 seconds to 120 seconds.The evaluation indicators used in this paper are sensitivity(Se),specificity(Sp)and equilibrium error rate(BER).Experimental results show that compared with other features,the sequential statistical features proposed in this paper have achieved better results..