Using Mel-Spectrograms and 2D-CNNs to Detect Murmurs in Variable Length Phonocardiograms

As part of the George B. Moody PhysioNet Challenge 2022, we developed a computational approach for identifying abnormal cardiac valve function from phonocardiograms (PCGs). Our team, uke-cardio, developed a deep learning model that uses melspectograms of up to four different auscultation locations. CutMix and smooth labels were used to simultaneously learn different tasks. On the hidden test set, the classifier achieved a weighted accuracy score of 0.735 for murmur detection (ranked 15 out of 40 teams) and a challenge cost score of 11990 (ranked 8 out of 39 teams) on the outcome detection task. The source code is available at: github.com/msknorr/cinc22-pcg