Beat-wise Uncertainty Learning for Murmur Detection in Heart Sounds

Abstract

This paper introduces a murmur detection solution (Team SeaCrying) to the PhysioNet Challenge 2022. The method is based on beat-wise uncertainty learning for heart sounds. The target task is to distinguish the present and absent state for murmur, with an outlier situation indicated as unknown in the challenge. Two uncertainties induced by outlier noise and fuzzy sounds are addressed while beat segmentation and murmur discrimination, respectively. In beat segmentation stage, we employ a confidence branch trained by a frame-level noise contrastive framework to quantify the uncertainty for out-ofdistribution episodes. Then we transmit the groups of five effective heart beats to the murmur discriminator and each beat is concatenated by a systole (containing S1 and S2) and a diastole. To alleviate the issue of disability for the model learning unknown sounds, we adopt an uncertainty estimation module on the basis of binary classification for murmur detection. The unknown samples will lead to a highly uncertainty score. As well a cross-beat decision strategy is designed for the same phonocardiogram recording in the final stage. Our contribution submitted for evaluation achieved a weighted accuracy of 0.601 on the test set in murmur detection and 0.837 for the outcome classification task. The cost metric score the test set is 12126 for murmur detection and 15083 for outcome classification.