Generative adversarial network augmented data for improved heart sound abnormality detection.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Cardiovascular Diagnostics

Background:

• The PhysioNet/Computing in Cardiology (CinC) Challenge 2016 dataset, crucial for automated heart sound analysis, suffers from limited size and class imbalance, particularly for coronary artery disease (CAD).
• Underrepresentation of CAD cases hinders the development of robust machine learning (ML) and deep learning (DL) models for accurate heart sound classification.
• Existing data augmentation techniques are insufficient to overcome the severe imbalance in medical audio datasets.

Purpose of the Study:

• To address the limitations of the CinC 2016 dataset by synthesizing realistic coronary artery disease (CAD) heart sound segments using generative adversarial networks (GANs).
• To augment the existing imbalanced dataset with high-quality synthetic heart sound data to improve the performance of classification models.
• To evaluate the effectiveness of GAN-based data augmentation compared to traditional methods for heart sound analysis.

Main Methods:

• Implementation of a Progressive Wasserstein Generative Adversarial Network (GAN) architecture to generate synthetic heart sound segments.
• Generation of CAD-like heart sound audio, focusing on capturing spectral and temporal characteristics.
• Assessment of synthetic audio quality using Fréchet Audio Distance (FAD) and application of novel post-processing steps like bandpass filtering.
• Augmentation of the imbalanced heart sound dataset with generated samples and evaluation of five classification models.

Main Results:

• The Progressive Wasserstein GAN successfully generated high-fidelity synthetic heart sound segments comparable to real CAD and healthy samples (FAD scores of 1.43 and 2.23, respectively).
• GAN-based data augmentation significantly improved the performance of five classification models, outperforming traditional augmentation and cost-sensitive learning methods.
• Augmented models demonstrated superior sensitivity, specificity, and precision in heart sound classification tasks.
• Post-processing steps like bandpass filtering further enhanced the quality and utility of the synthetic audio data.

Conclusions:

• Generative adversarial networks (GANs) offer a powerful and scalable solution for addressing data scarcity and class imbalance in medical audio datasets.
• GAN-based data augmentation substantially enhances the generalizability and robustness of heart sound classification models.
• This approach provides a cost-effective method for improving diagnostic tools in biomedical signal processing, particularly for conditions like coronary artery disease.