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

Erika Bondareva, Jing Han, Tong Xia, Cecilia Mascolo
University of Cambridge


Abstract

Auscultation, the process of using a stethoscope for diagnostics, is a challenging task for medical professionals and requires years of training. As a result, the field of automated auscultation has been growing in popularity in the past decade. Previous efforts in the field focused on achieving high accuracy, with confident, albeit sometimes wrong, classifiers. Such model over-confidence is especially dangerous in healthcare 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. In order to effectively 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, a support vector machine used spectral features for identification of unknown samples. We then used a Deep Neural Network (DNN) with a set of hand-crafted audio features 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. With our approach we achieved 63% and 69% sensitivity and specificity of murmur, scoring 0.519 and 11301 during the challenge for murmur and outcome prediction tasks, respectively.