Exploring Wav2vec 2.0 Model for Heart Sound Analysis

This study investigates the utilization of the wav2vec 2.0 framework for heart sound analysis, with the intention to reduce the reliance on supervised data provided by cardiologists. Leveraging the self-supervised learning concept, wav2vec learns audio context representations by solving pretext tasks, thus diminishing the need for labeled data in downstream tasks. The model was pre-trained on heart sounds, lung sounds, and hospital ambient noise to sufficiently grasp the acoustic features of biological sounds. To tailor this pre-trained model for heart sound analysis, we employ either freezing the CNN encoder or the entire model to function as a feature extractor, followed by training the classifier heads using LSTM with self-attention framework. This framework is evaluated for heart murmur detection or segmentation of S1 and S2 sounds. Even though our pre-trained model was trained on a smaller dataset with 52 hours of audio, its performance is slightly better than the model pre-trained on large corpus of 960 hours for speech representation. The best-weighted accuracy scores achieved are 81% for heart murmur detection and 97.56% for S1 and S2 sound segmentation, respectively.