Heart Murmur Classification Using Transfer Learning and Snapshot Ensemble Method

Automated heart sound classification is crucial for diagnosing cardiovascular diseases (CVDs). The convergence of medical big data and advanced artificial intelligence has intensified efforts to enhance heart sound classification, especially through the development of sophisticated deep learning methods. This study leverages deep learning techniques to autonomously and precisely detect heart murmurs from phonocardiogram (PCG) recordings. Our proposed system undergoes pre-processing steps such as denoising, chunking, and padding, along with extracting mel-spectrograms. These features are subsequently input into pre-trained Imagenet models, employing the snapshot ensemble method to enhance system performance. We assess the effectiveness of this approach using the CirCor Digiscope 2022 dataset, a contemporary public PCG dataset sourced from the reputable Physionet online database. Experimentally, our approach achieves the highest performance with transfer learning from the pretrained ResNet, DenseNet and MobileNet models. In addition, snapshot ensemble method of ResNet and DenseNet models also proves its efficiency to the heart murmur sound classification task.