Machine-Learning-based Aortic Stenosis Detection for Electronic Stethoscope

Abstract

Analyzing the phonocardiogram (PCG) collected by a electronic stethoscope can help to quickly diagnose structural heart diseases. In this paper, we aim at automatic detection of aortic stenosis (AS) based on PCG, with the aid of two proposed machine-learning based methods. The first method is a model-dependent method, a Gaussian mixture model - hidden Markov model (GMM-HMM) method, which exploits the temporal relationship among states in cardiac sounds. The second one is a data-dependent method, implemented by a 1D/2D-fused-feature-based convolutional neural network. The results of comparative experiments showed that both methods can fulfill the AS automatic diagnosis task to a certain extent, and in most cases CNN scored higher than GMM-HMM, which indicated the importance of automatically learning an unknown model from data in this problem, although the GMM-HMM method with fewer parameters also have potential advantages in practice.