A Heart Sound Classification Method Based on Time Series Analysis

Abstract

Auscultation of heart sound is the main diagnostic method of cardiovascular and cerebrovascular diseases. However, the traditional heart auscultation relies too much on the sensitivity of human ear and the subjective experience of doctors, which makes it difficult to make a correct judgment of heart sound. This paper proposes a heart sound signal classification method based on time series. The use of advanced signal processing methods and deep learning methods can effectively alleviate this problem. The method first uses the biorthogonal wavelet base to denoise, and uses band-pass filtering to filter out the unqualified frequency band signals. By calculating the wavelet entropy range of all heart sound data, it is used to filter out the fuzzy heart sound data that is not within the threshold range; Then, according to the contribution of each feature's SHAPLEY value to the model, the MFCC feature combination that is most suitable for the model is selected; Finally, a TCN-LSTM model is designed to process timing information. Experiments show that this method can accurately detect the benign and malignant of audio data.