Study of Machine Learning Algorithm on Phonocardiogram Signals for Detecting of Coronary Artery Disease

Abstract

Several methods of detecting coronary artery disease (CAD) have been developed, but they are expensive and generally use an invasive catheterization method. This research provides a solution to this problem by developing an inexpensive and non-invasive digital stethoscope for detecting CAD. To prove the effectiveness of this device, twenty-one subjects consisting of 11 CAD patients and 10 healthy people from Hasan Sadikin Hospital Bandung were selected as validation test participants. In addition, auscultation was carried out at four different locations around their chests, such as the aorta, pulmonary, tricuspid, and mitral. Then the phonocardiogram data taken from the stethoscope were analyzed using machine learning. To obtain optimal detection accuracy, several types of kernels such as radial basis function kernel (RBF), polynomial kernel and linear kernel of Support Vector Machine (SVM) have been analyzed. The experimental results show that the linear kernel outperforms compared to others; it provides a detection accuracy around 66%. Followed by RBF is 56% and Polynomial is 46%. In addition, the observation of phonocardiogram signals around the aorta is highly correlated with CAD, giving an average detection accuracy for the kernel of 66%; followed by 44% tricuspid and 43% pulmonary.