Phonocardiographic Murmur Detection by Scattering-Recurrent Networks

We describe an automatic detector of phonocardiogram murmurs. Our detector composes the scattering transform (ST) and a long short-term memory (LSTM) network. It is trained on data as part of the Heart Murmur Detection from Phonocardiogram Recordings: The George B. Moody PhysioNet Challenge 2022. The ST captures short-term temporal ECG modulations while reducing its sampling rate to a few samples per typical heart beat. We pass the output of the ST to a depthwise-separable convolution layer which transforms responses separately for each ST coefficient and then combines resulting values across ST coefficients. At a deeper level, 2 LSTM layers integrate local variations of the input over long time scales. We train in an end-to-end fashion as a classification problem with three murmur classes: present, absent or unknown. Additionally, we use the model to classify clinical outcome as normal or abnormal. These two classifications determine whether clinical followup should occur. Our team “PAWPCG” obtained an official score on the hidden test data of 0.637 for weighted accuracy on murmur classification (rank: 27 of 40 teams) and a clinical outcome cost of 15083 (rank: 32 of 39 teams).