Detection of Murmurs from Heart Sound Recordings with Deep Residual Networks

Abstract

Cardiac auscultation is an effective method to screen hemodynamic abnormalities. As part of the George B. Moody PhysioNet Challenge 2022, this paper aims to propose an automated algorithm to identify the presence of murmurs in heart sounds from multiple auscultation locations and to determine whether the heart sounds signal is normal. Two methods are explored. In method one, we perform a series of pre-processing such as denoising and segmentation of the heart sounds signal, extract Log Mel-Spectrogram as features, and use fastai's built-in xResNet 18 pre-trained model for classification. In method two, we extract Mel-frequency cepstral coefficients (MFCCs) as features without any pre-processing and build a customized model based on deep residual networks using one-dimensional convolutional neural layers. Our team, USST_ Med, received a challenging score of weighted accuracy of 0.642 (ranked 26th out of 40 teams) and cost of 14529 (ranked 30th out of 39 teams) on the final hidden test set.