Heart Murmur Detection Using Ensemble of Deep Learning Classifiers for Phonocardiograms Recorded from Multiple Auscultation Locations

Abstract

A digital phonocardiogram (PCG) provides an opportunity for automated screening in resource-constrained environments. As part of the George B. Moody PhysioNet Challenge 2022, our team, Life_Is _Now, developed a computational approach using an ensemble of deep learning classifiers for identifying abnormal cardiac function from PCG. A stratified 5-fold cross-validation was used for model development and evaluation for murmur and clinical outcome identification. The backbone of our trained classifiers is a modified pre-trained deep convolutional neural network on AudioSet-Youtube corpus (YAMNet) and transfer learning. The YAMNet model is modified and finetuned on the publicly available PhysioNet dataset. Our murmur and clinical outcome classifiers received a weighted accuracy score of 0.831 and a Challenge cost score of 14,850 from cross-validation on the public training set. Our murmur scores were 0.678 and outcome score were 10,518 on the hidden validation set. However, we did not receive the official score for the hidden test set as our entry crashed in evaluation on the test set.