Murmur Detection and Clinical Outcome Classification Using a VGG-like Network and Combined Time-Frequency Representations of PCG Signals

Abstract

For the George B. Moody PhysioNet Challenge 2022, our team, PhysioDreamfly, developed a deep neural network approach for detecting murmurs and identifying abnormal clinical outcomes from phonocardiograms (PCGs). In our approach, a VGG-like CNN model is used as the classifier. Images consisting of Log-Mel spectrograms and wavelet scalogram that transformed from unsegmented PCGs are used as model inputs. We combined the murmur and outcome labels to address the two tasks as one multi-label task, and introduced a weighted focal loss function to optimize the model. Our murmur detection classifier received a weighted accuracy score of 0.752 (ranked 11th out of 40 teams) and Challenge cost score of 12831(ranked 18th out of 39 teams) on the hidden test set.