Exploring wav2vec 2.0 model for heart murmur detection

The lack of access to cardiology resources in many regions of the world has motivated the development of automatic diagnostic systems based on cardiac signals. In recent years, a wide range of supervised learning models have been proposed that can make an initial diagnosis of heart disease from heart sounds. To achieve high accuracy, however, such supervised learning models generally require a large amount of labeled data, which can be costly to obtain. In this regard, self-supervised learning has been recently employed to reduce the over-reliance on annotated data. Wav2vec 2.0 is an audio self-supervised learning model that has shown promising results in a variety of speech-related tasks. In this paper, we adapted the wav2vec 2.0 for murmur detection from heart sound signals. For this purpose, we pre-trained and fine-tuned this model on the Circor DigiScope heart sound dataset. The results confirm the feasibility of using the wav2vec 2.0 model for heart sound classification. The model shows a competitive performance by achieving a weighted accuracy of 0.80 and a UAR of 0.70 for murmur detection on the holdout test set. To investigate the impact of the fine-tuning data size on the downstream performance, we also fine-tuned the wav2vec 2.0 model on small sizes of annotated data. The results confirm that this model is robust to small fine-tuning data sizes, and as a result, can reduce our reliance on large, annotated heart sound data.