Classification of heart murmurs using an ensemble of residual CNNs

Petr Nejedly1, Jan Pavlus1, Radovan Smisek2, Zuzana Koscova3, Eniko Vargova3, Ivo Viscor3, Pavel Jurak3, Filip Plesinger3
1Institute of Scientific Instruments of the Czech Academy of Science, 2Brno University of Technology, Faculty of Electrical Engineering and Communication, Department of Biomedical Engineering, 3Institute of Scientific Instruments of the CAS


This paper introduces a solution (team ISIBrno-AIMT) to the unofficial round of PhysioNet Challenge 2022. The main goal of the challenge was a classification of heart murmurs from phonocardiographic recordings into three mutually exclusive classes (i.e., present, unknown, and not present). We introduced a classification method based on the ResNet architecture with a multi-head attention mechanism. The model is optimized using a mixture of loss functions, i.e., binary cross-entropy and custom challenge score loss function that provides a differentiable approximation of challenge score. The objective challenge function was generally designed to reflect the cost of algorithmic pre-screening, expert screening, treatment, and missed diagnoses. Our proposed solution was ranked 1st (out of 166 submissions) in the unofficial challenge round, achieving a challenge score of 464. The cost of general practitioner screening of all patients in the publicly available dataset is approximately 540. Therefore our solution might be potentially cost-saving.