Towards Uncertainty-Aware Murmur Detection in Heart Sounds via Tandem Learning

Abstract

The field of automated auscultation has been growing in popularity in the past decade due to manual auscultation being a challenging task requiring years of training. Many efforts in the field focus on achieving high accuracy, with confident, albeit sometimes wrong, classifiers. Such model over-confidence is especially dangerous in health-care setting. Leveraging the release of the new heart sound dataset as a part of PhysioNet 2022 challenge, we explored a novel murmur detection methodology using uncertainty-aware tandem learning. To separate unknown samples and detect heart sounds with murmur present, we developed two binary classifiers, under the assumption that training two models to solve simpler tasks could improve the overall sensitivity. First, we used a support vector machine for identification of unknown samples, followed by a Deep Neural Network (DNN) for prediction of murmur. In addition, we implemented uncertainty estimation in DNN using Monte Carlo dropouts for further eliminating any samples that should be labelled as unknown. Our team mobihealth achieved 63% and 69% sensitivity and specificity of murmur, scoring 0.467 (ranked 34th out of 40) and 11032 (ranked 25th out of 39) on the hidden validation set and 0.374 (ranked 40th out of 40) and 18754 (ranked 39th out of 39) on the hidden testing set during the challenge for murmur and outcome prediction tasks, respectively.