Generative artificial intelligence for heart sound analysis: From model architectures to clinical applications

Abstract

Cardiac auscultation involves listening to heart sounds to diagnose cardiac conditions. Accurate interpretation of these sounds requires considerable experience, which limits consistency among clinicians and across settings. Although deep learning offers a path toward automation, data scarcity, class imbalance, and signal complexity pose challenges for its application to heart sound analysis. This study presents a pioneering comprehensive examination of how generative artificial intelligence (generative AI) models, including generative adversarial networks, variational autoencoders, diffusion models, and transformer-based architectures, can address key challenges in heart sound analysis. Specifically, these models mitigate core limitations by synthesizing high-fidelity heart sound signals for data augmentation, improving signal quality through advanced denoising, and enabling semi-supervised and unsupervised detection of rare pathological conditions. In addition, we discuss their roles in advanced feature learning and cross-modal translation. Despite notable progress, we highlight future challenges, including ensuring physiological fidelity, enhancing model interpretability, and validating clinical utility. In conclusion, generative AI addresses a key translational challenge by enabling accessible, accurate, and scalable expert cardiac evaluation, thereby supporting the implementation of precision cardiology.