PCG Frame Classification by Classical Machine Learning Methods Using Spectral Features and MFCC Based Features

Abstract

Cardiovascular diseases are some of the most common diseases today. Congenital abnormalities, diseases caused by impaired heart rhythm, vascular occlusion, post-operation arrhythmias, heart attacks and irregularities in heart valves are some of the various cardiovascular diseases. Early recognition of them is very important for obtaining positive results in treatment. For this purpose, it is tried to diagnose and detect cardiovascular diseases by listening to the sounds coming from the heart. During the rhythmic work of the heart, the contraction and relaxation of the heart chambers and the filling and discharge of blood from the heart into the veins create the sounds that are identified with the heart. Among the characteristic sounds of the heart, there can be some sounds similar to rustling which are indicators of pathological conditions. These unexpected sounds, similar to rustling, are called heart murmurs. Phonocardiograph device is used to record these mechanical sounds via microphone. Heart sounds recordings captured by a phonocardiograph device are called phonocardiograms (PCGs). Expert physicians try to detect the heart murmurs by listening to the heart sounds and examining PCGs. Ambient noise, the squeak of the microphone, and the patient's breathing sounds are the factors that make this task more difficult and challenging. Computer-aided systems supported with machine learning, signal processing and artificial intelligence algorithms offer solutions to help physicians in this regard. In this study, detection of heart murmur from PCG frames was examined. PCG frames of equal length, obtained by fragmenting the PCG recordings into 1-second-long frames, were classified by widely used machine learning methods namely C4.5 decision tree, Naive Bayes, Support Vector Machines and k-nearest neighbor. To train those classifiers we used spectral features of PCG signals, averages of MFCC values and some refined features obtained from a deep learning model which was inputted MFCC values. At the end of this manuscript the accuracies of those machine learning methods were compared.