We present a deep learning model for the automatic detection of murmurs and other cardiac abnormalities from the analysis of digital recordings of cardiac auscultations. This approach was developed in the context of the George B. Moody PhysioNet Challenge 2022.
More precisely, we consider multi-objective neural networks, with several Transformer blocks at their core, trained to perform 3 distinct tasks simultaneously: murmur detection, outcome classification and audio signal segmentation. We also perform pre-training with the 2016's Challenge data.
We entered the challenge under the team name matLisboa. Our (best) results on the hidden validation dataset (public Challenge leaderboard) were:
Murmur score (weighted accuracy): 0.754.
Outcomes score (cost): 9512.