Heart Murmur Classification in Phonocardiogram Representations Using Convolutional Neural Networks

Abstract

Heart murmurs are sounds made by rapid blood flow in the heart. Abnormal heart murmurs can be a sign of serious heart conditions such as arrhythmia and cardiovascular diseases. Therefore, heart murmur classification is crucial for early detection of such conditions. To this end, we study the heart murmur classification problem training selected convolutional neural network (CNN) models (such as VGGNet and ResNet) using various signal representations (such as spectrogram, mel-frequency cepstral coefficient (MFCC), and shorttime Fourier transform (STFT)) of the phonocardiograms in the public PASCAL CHSC dataset. Our preliminary results show that ResNet outperforms VGGNet across all metrics and representations, consistent with the recent published works we can find in literature. Unlike some of these works, however, we see MFCC and STFT in general more effective with higher test accuracies than spectrogram across all CNN models. Looking forward, we propose to study other effective models (such as InceptionV3 and Vision Transformer) to predict heart murmur conditions in phonocardiogram representations including spectrogram, MFCC and STFT, as well as others like Wigner Ville distribution.