Research On Quality Evaluation And Classification Of Heart Sound

Cardiovascular disease has become the biggest threat to contemporary human health.As a physiological signal that is non-invasively collected on the body surface and is rich in cardiovascular pathological information,heart sound has important reference value in the auxiliary diagnosis of cardiovascular diseases.To study the automatic classification algorithm of heart sound signals,this thesis uses heart sound as the data basis,based on the quality evaluation algorithm and machine learning models.It aims to identify disease information from the short-term record to assist clinical cardiovascular disease diagnosis.This thesis uses the heart sound data of 2575 normal people and 665 patients provided by the 2016 Cardiology Challenge,and mainly conducts the following research:(1)Research on heart sound signal quality evaluation based on the sliding window and wavelet decomposition features.For heart sound signals that are susceptible to interference from environmental noise,a sliding window is used to simulate the auscultation process of clinicians.The root mean square of successive differences and the ratio of zero crossings are used as evaluation indicators,and the heart sound signal quality is evaluated and screened through preset indicator thresholds to provide a data basis for subsequent experiments.(2)Classification algorithm of heart sound based on Mel cepstrum features and wavelet scattering features.Two feature extraction methods and heart sound classification algorithms without waveform segmentation and heart beat labeling were studied.Feature extraction of Mel cepstrum:Mel cepstrum converted from Mel frequency standard based on the human ear hearing mechanism are used as classification features.Five typical machine learning classification models including K-nearest neighbor,support vector machine,decision tree,AdaBoost algorithm and convolutional neural network are studied and compared.Feature extraction of wavelet scattering:Based on the translation invariance of the wavelet scattering transform,the wavelet scattering network is constructed,and the scattering coefficients output layer by layer of the wavelet scattering network is regarded as the classification feature.The effects of window and indicators parameters on classification performance in the quality assessment are further analyzed and the effect differences between feature fusion and result fusion are compared.The model’s sensitivity,specificity,accuracy,and the mean of accuracy were 96.62%,90.65%,92.23%,and 93.64%,respectively.It is proved that the method presented in this paper has excellent classification ability of heart sound signal,and has a particularly important reference value in the realization of auxiliary diagnosis of cardiovascular diseases.