Segmentation-free Heart Murmur Detection Using Deep Learning

Ľubomír Antoni1, Erik Bruoth1, Alexander Szabari1, Gabriela Vozáriková1, Peter Bugata2, Peter Bugata Jr.2, David Gajdos2, Dávid Hudák2, Vladimíra Kmečová2, Monika Staňková2
1Pavol Jozef Šafárik University, 2VSL Software, a.s.


Abstract

The Objective of The George B. Moody's PhysioNet Challenge 2022 was to develop an algorithm that detects the presence or absence of murmurs in patients from multiple heart sound recordings captured in multiple auscultation locations. Our proposed model is based on a one-dimensional residual convolutional neural network, which processes short slices of the individual raw non-segmented phonocardiogram (PCG) recordings. The model addresses binary classification, where the negative class includes all PCGs from patients who did not have a murmur and the positive class (suspicious PCGs) contains all other PCG recordings. The overall predictions are obtained as the maximum probability of the positive class from multiple, evenly distributed slices of each of the recordings. In the unofficial round, we used a simple decision rule: a patient is considered as Positive and sent to a general practitioner if any of his PCG recordings are classified as positive, otherwise we classify him as Negative. Our team CeZIS obtained a Challenge score 466 and placed 2nd. Possible improvements in the official round could include more complicated solutions using spectrograms, different data augmentation techniques and adding data from another publicly available PCG datasets.