Multi-Task Prediction of Murmur and Outcome from Heart Sound Recordings

As part of the George B. Moody PhysioNet Challenge 2022, our team, prna, developed a novel approach to automatically detect the presence of heart murmurs and predict abornomal clinical outcomes from the heart sound recordings. To train the model, each heart sound recording is first divided into multiple 3-second segments. Then each segment is transformed into a time-domain embedding vector through a convolutional neural network (CNN). In parallel, the Mel-frequency cepstrum (MFCC) representation of the segment is transformed into a frequency-domain embedding vector using CNN. These embedding vectors and the demographic variables are concatenated and then used as input to two separate networks built to predict the presence of heart murmurs and clinical outcomes respectively. The network parameters are optimized to jointly predict both targets using multi-task learning. Our murmur detection classifier received a weighted accuracy score of 0.694 (ranked 20th out of 40 teams) and Challenge cost score of 11,403 (ranked 2nd out of 39 teams) on the hidden test set.