Detecting Heart Murmur using a Modular Model

Mohammadreza Heydarian1, Ali Shakiba1, Meraj Hashemizadeh2
1Department of Computer Science, Vali-e-Asr University of Rafsanjan, 2Mila & DIRO, Université de Montréal


Introduction: In the last two decades, heart disease has been the leading cause of death globally. Although an innocent heart murmur does not need treatment, an abnormal heart murmur can be caused by heart problems such as valvular heart disease. Early detection of heart murmur and finding its cause could help patients prevent further issues. The 2022 PhysioNet challenge offers the opportunity to develop AI models for heart murmur detection. We (the Veteran team) propose a wide and deep CNN modular framework to detect heart murmur from phonocardiogram signals for the PhysioNet/CinC Challenge 2022. Methods: Each phonocardiogram signal was down-sampled to 1000 Hz and segmented to 5-second length with windowing longer signals with a 3-second overlap sliding window. A multi-layer one-dimensional deep CNN (1D-dCNN) module was used to extract sequentially related features embedded in the signals automatically. In parallel to the 1D-dCNN module, a shallow neural network was added to employ manually extracted features from meta-data and signals. These two parallel modules were concatenated and fed to a classifier module, creating a ‘wide and deep modular framework.’ Our proposed model allows adding/removing modules for optimizing the classifier. The model was trained until stable validation loss was achieved over 5 Epochs. Results: Our method achieved a cross-validated scoring metric (accuracy) of 0.8288 on training data. The challenge validation score provided by the PhysioNet challenge on a hidden dataset was 1763. Conclusion: Our proposed framework offers modularity and demonstrates potential with its wide module. While excluding some meta-data such as age and weight improved our validation score from 1882 to 1763, extracting the murmur-related features of the signal could increase the performance of our classifier. The proposed model provides the ability to integrate heart murmur extracted features from phonocardiogram recordings in future modules to improve the overall performance.