Multitask and Transfer Learning for Murmur Detection in Heart Sounds

João Costa1 and Rui Rodrigues2
1University Institute of Lisbon - ISCTE and CAMGSD - IST/ULisboa, 2DM-FCT NOVA


The detection of murmurs via cardiac auscultation provides critical insights into heart malfunctioning. The design of algorithms for the detection of such pathology can provide valuable information in situations where there is difficulty in accessing health services. In this context, we present our contribution for the George B. Moody PhysioNet Challenge 2022.


Motivated by both the need to compress the size of the original data and by the fact that heart murmurs emanate from turbulent flows, which have a characteristic profile in frequency space, we start by applying the Short Time Fourier Transform to the audio signals and truncate the corresponding spectrum at frequencies smaller than 1000 Hz.

We use a neural network to classify each heart sound signal as "positive", "unknown" or "negative" for the presence of murmurs. This information is then aggregated to produce a similar classification at the level of each patient. Our model is trained to perform two simultaneous tasks: a principal task corresponding to the aforementioned classification and a secondary task that attempts to obtain the segmentation of the original signals into S1, Systolic, S2 and Diastolic waves. To perform this multitask learning our network starts with three convolution blocks, followed by a first transformer encoder block. Then our model splits into two branches, one for each task:

The entire model was pretrained with 2016 CinC/Physionet Challenge Data and then fine-tuned on the 2022 Challenge data.