A Comparative Study of Heart Sound Signal Classification Based on Temporal, Spectral and Geometric Features

This paper presents a study of murmur recognition in heart sound signal with different machine learning algorithms using combined features. The heart sound signal is required to be pre-processed and segmented first before performing a feature generation and classification process. In this work, two hundred segmented heart cycle samples were parameterized based on temporal, spectral and geometric features as input of cubic Support Vector Machine (SVM), quadratic Discriminant Analysis (DA) and fine K-Nearest Neighbour (KNN) classifiers. The significant of this study is to support the hypothesis in statistic that said higher number of samples provide more accuracy of classifier and smaller margin error. To determine which classifier algorithm is outperformed, cross-validation technique is conducted and varied with different number of samples. The findings indicate cubic Support Vector Machine (SVM) classifier model yields the best performance, achieving an accuracy of 0.97 and F1-score of 0.98 at higher number of samples.