Heart Murmur Detection from Phonocardiogram Based on Residual Neural Network with Classes Distinguished Focal Loss

Abstract

The George B. Moody PhysioNet Challenge 2022 focused on detecting the presence or absence of murmurs from multiple auscultation locations heart sound recordings. Our team, MetaHeart, proposed a novel approach to detect heart murmurs by combing residual neural network and class distinguished focal loss. Firstly, the phonocardiogram (PCG) recordings were converted to the time-frequency Mel spectrograms to obtain a richer representation of cardiac mechanical activity. Secondly, a modified residual neural network with 30 layers was designed to extract the complex pathological patterns of heart murmurs. Thirdly, a joint loss combing class distinguished focal loss and center loss was designed in our approach. The class distinguished focal loss can give different degrees of attention to misclassified samples from different categories. The center loss learns a center for deep features of each class. The 15 seconds recordings from five corresponding auscultation locations were preprocessed and concatenated as model inputs for end-to-end training, and the prediction probabilities of 3 murmurs categories or 2 outcome categories were outputs. Finally, our murmur detection classifier received a weighted accuracy score of 0.72 (ranked 18th out of 40 teams) and Challenge cost score of 12536 (ranked 15th out of 39 teams) on the hidden test set.