Heart Murmur Detection Using Ensemble of Deep Learning Classifiers for Phonocardiograms Recorded from Multiple Auscultation Locations

Saman Parvaneh, Zaniar Ardalan, Joomyung Song, Kathan Vyas, Cristhian Potes
Edwards Lifesciences


Background: The George B. Moody PhysioNet Challenge 2022 aims to encourage algorithm development for pre-screening of abnormal heart function using Phonocardiogram (PCG) recorded from multiple auscultation locations. This article proposes a modified pre-trained convolutional neural network (CNN) to classify PCGs recorded from four auscultation locations for a heart murmur and outcome detection. Method: Data was first segmented into windows of 0.98s with 50% overlap and 16kHz. Raw audios convert to Mel spectrogram with windows length of 25ms, an overlap of 15ms, and 64 frequency bands covering the range 125-7500 Hz. Furthermore, the murmur majority class (i.e., absent) was randomly undersampled by factor four. A stratified 5-fold cross-validation was utilized for model development. Pre-trained CNN on AudioSet-Youtube corpus (YAMNet) is modified and finetuned on the PhysionNet dataset. For the murmur classification, the last three layers of YAMNet were removed and replaced for 3-class classification. Furthermore, the loss function was weighted 5, 3, and 1 for present, unknow, and absent classes. The same approach was followed for outcome classification except for the output layer adopted for 2-class classification, normal and abnormal. The modified YAMNet was trained on both 2016 and 2022 PhysioNet challenge data to produce two models. Then the top performing model of Physionet/Computing in Cardiology Challenge 2016, an ensemble of AdaBoost and CNN, was slightly modified based on the challenge data. Then, maximum voting on the output of three models was used for outcome identification. Results: The best challenge score achieved by our team (Life_Is_Now) for the hidden validation dataset was 0.678 on murmur detection (ranked 26th out of 62 teams) and 10,518 on outcome detection (ranked 34th out of 62 teams). Discussion: The result of our approach is promising. We plan to improve the algorithm's performance by denoising the data and data augmentation.