PACNet-HS: A Physics-Aware Contrastive Network for Heart Sound Classification

Heart sound analysis has emerged as a promising non-invasive and cost-effective method for early cardiac screening, particularly in resource-constrained settings. However, current heart sound classification methods often rely on image-based adaptations or overlook the sequential nature of heart sound signals. In this study, we propose Physics-Aware Contrastive Network for Heart Sound (PACNet-HS). The model introduces a dual-branch contrastive learning framework that processes adjacent cardiac cycles using a shared convolutional neural network and bidirectional long short-term memory encoder and a contrastive feature fusion classifier. To enhance prediction stability and representation consistency, we design a physics-inspired loss function that includes classification, feature similarity, prediction agreement, and regularization terms. Experiments conducted on the PhysioNet/CinC Challenge 2016 dataset demonstrate that PACNet-HS achieves competitive performance while maintaining low model complexity. Compared to existing baseline models, our approach shows improved robustness and better generalization across noisy and imbalanced datasets, offering practical value for portable and real-time auscultation systems.