Generative deep learning and signal processing for data augmentation of cardiac auscultation signals: Improving model robustness using synthetic audio

Accurately interpreting cardiac auscultation signals is essential for diagnosing and managing cardiovascular diseases. However, the paucity of labelled data inhibits classification models’ training. Researchers have turned to generative deep learning techniques alongside signal processing to augment existing data and improve cardiac auscultation classification models. However, the primary focus of prior studies has been on model performance rather than robustness. Robustness, in this case, is defined as both in-distribution and out-of-distribution performance by measures such as Matthew’s correlation coefficient. One contribution of this work is to show that more robust abnormal heart sound classifiers can be trained using an augmented dataset. The augmented dataset includes both signal processing techniques and synthetic phonocardiograms conditionally generated using the WaveGrad and DiffWave diffusion models an approach that, to the best of our knowledge, is the first work of its kind. The efficacy of the proposed data augmentation approach is evaluated on an example convolutional neural network, trained on the original and augmented data. It is found that both the in-distribution and out-of-distribution performance can be improved over various datasets for neural networks trained with this augmented dataset. Results show significant performance improvements. Specifically, in-distribution accuracy, balanced accuracy, and Matthew’s correlation coefficient (MCC) increased by 2.5%, 4.1%, and 0.066, respectively. The greatest out-of-distribution improvements were observed on one dataset, where accuracy, balanced accuracy, and MCC increased by 43.1%, 20.2%, and 0.297, respectively. These improvements across all metrics highlight that augmented datasets significantly address issues of imbalanced data, ultimately leading to more generalisable and robust classifiers. • Diffusion model developed to generate PCG signals from ECG signals. • Enables use of large ECG datasets by creating corresponding PCG data. • First known application of diffusion models for PCG signal synthesis. • Advanced augmentation methods applied to PCG and ECG signals. • Improved the in-distribution performance by: 2.5% acc, 4.1% bal. acc. • Improved the out-of-distribution performance by up to: 43.1% acc, 20.2% bal. acc.