PCGmix: A Data-Augmentation Method for Heart-Sound Classification.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiovascular diseases (CVDs) are a primary cause of global mortality, necessitating early risk identification for effective prevention.
• Deep learning models show promise for automated CVD detection using phonocardiogram (PCG) data, but require substantial datasets.
• Data augmentation is a key strategy to address data scarcity in deep learning applications.

Purpose of the Study:

• To introduce PCGmix, a novel data augmentation technique specifically designed for phonocardiogram (PCG) data.
• To evaluate the effectiveness of PCGmix in enhancing deep learning models for cardiovascular disease (CVD) detection, particularly in data-limited scenarios.

Main Methods:

• Developed the PCGmix algorithm, which segments and reassembles PCG recordings with precise interpolation to preserve diagnostic features.
• Applied PCGmix to a public dataset of normal and abnormal heart sound recordings.
• Conducted experiments with varying dataset sizes to assess the impact of augmentation on model performance compared to state-of-the-art methods.

Main Results:

• PCGmix outperformed existing time-series data augmentation techniques.
• In limited data settings, PCGmix achieved accuracy comparable to models trained on significantly larger datasets (31% to 69% more data).
• The proposed method demonstrates superior performance across different dataset sizes.

Conclusions:

• PCGmix is an effective data augmentation technique for phonocardiogram (PCG) data.
• The method significantly improves the accuracy of deep learning models for cardiovascular disease (CVD) detection.
• PCGmix offers a valuable solution for improving CVD detection in data-constrained environments.