Maiby's Algorithm: A Two-Stage Deep Learning Approach for Murmur Detection in Mel Spectrograms for Automatic Auscultation of Congenital Heart Disease

Abstract

Congenital heart disease (CHD) is a major cause of death for newborns, especially in low resources countries, due to limited access to heart specialists for timely diagnosis. As part of the George B. Moody PhysioNet Challenge 2022, we propose an automatic algorithm to detect CHD murmurs from digitally recorded heart sounds annotated by specialists. To train and validate our model, we use a dataset with 5282 heart sounds collected from 1568 children in the Paraiba state of Brazil recorded from multiple auscultation locations. Our team, named One_Heart_Health, used a two-stage strategy that combines embeddings from Mel spectrograms generated from audio segments and a final classifier that combine those embeddings to deliver the final classification per individual. On the official hidden test, we reached a weighted accuracy score of 0.729 (ranked 17th out of 40) and a challenge cost score of 13283 (ranked 23th out of 39). In our internal 5-fold cross-validation experiments, our approach reached a sensitivity of 0.76 ± 0.10 and a specificity of 0.85 ± 0.11. We have shown that a deep learning approach for murmur detection has the potential to mimic heart specialists to provide timely identification of CHD.