An Improved Unsegmented Phonocardiogram Classification Using Nonlinear Time Scattering Features

Abstract

Abstract Phonocardiogram (PCG) signals highlight the relevant characteristics for the prediction of heart diseases or heart-related disorders. However, it is challenging to classify heart abnormality relying on an unbalanced PCG dataset due to low classification performance. Recently, several studies have attempted to predict heart abnormality based on segmented and unsegmented features extracted using PCG signals. This study aims to develop an automated PCG classification model eliminating any segmentation of the heart sound signal for predicting heart abnormality. So, we have proposed a new approach based on wavelet scattering transform to predict two classes of PCG signals, namely, normal and abnormal. Based on the wavelet scattering transform, five scattering time window features were extracted from each PCG signal. The PhysioNet 2016 PCG database has been used here to evaluate and compare the classification performance based on the k Nearest Neighbors (KNN) classifier. The proposed architecture used a KNN classifier with different distance functions (Euclidean, Cityblock, Chebyshev, Minkowsky, Correlation, Spearman and Cosine) and has been compared with other traditional classifiers (classification tree, linear discriminant analysis, support vector machine and ensemble). The proposed framework using nonlinear wavelet scattering features with a KNN classifier based Cityblock distance function achieved classification performance over the total datasets with accuracy, sensitivity and specificity values of 97.82%, 95.04% and 98.72%, respectively.