Generative adversarial network augmented data for improved heart sound abnormality detection

The PhysioNet/Computing in Cardiology (CinC) Challenge 2016 dataset has driven significant advancements in automated heart sound analysis using machine learning (ML) and deep learning (DL). However, these efforts are constrained by the dataset's limited size and severe class imbalance, particularly the underrepresentation of coronary artery disease (CAD) cases. This study addresses these limitations by employing generative adversarial networks (GANs) to synthesize realistic CAD-like heart sound segments, augmenting existing datasets to improve classification performance. A Progressive Wasserstein GAN architecture was implemented to generate high-quality audio segments that accurately capture CAD heart sounds' spectral and temporal characteristics. The quality of synthetic audio was assessed using the Fréchet Audio Distance (FAD), achieving scores of 1.43 and 2.23 when compared to reference CAD and healthy samples, respectively. Novel post-processing steps, including bandpass filtering, further enhanced the fidelity of the synthetic samples. By augmenting the imbalanced heart sound dataset with these samples, we observed substantial improvements in the performance of five classification models. The GAN-augmented training set outperformed traditional augmentation and cost-sensitive learning methods, demonstrating superior sensitivity, specificity, and precision. This study highlights the potential of GAN-based data augmentation to address critical challenges in medical audio datasets. It offers a scalable and cost-effective solution for improving the generalizability and robustness of heart sound classification models, paving the way for enhanced diagnostic tools in biomedical signal processing. • Progressive GANs synthesize high-quality heart sounds resembling CAD features. • Augmented datasets improve CAD heart sound classification performance. • Novel use of Progressive Wasserstein GANs for abnormal heart sounds synthesis. • Synthetic audios achieve FAD scores of 1.43 (CAD) and 2.23 (healthy). • Advanced GAN mechanisms ensure stable training and variety in synthetic samples.