Weight Light, Hear Right: Heart Sound Classification with a Low-Complexity Model

Cardiovascular diseases (CVDs) remain a leading cause of mortality, necessitating early self-diagnosis for effective management. Computer-aided CVD diagnosis via computerised auscultation has gained traction, advancing intelligent diagnostics. However, integrating complex neural networks into medical edge devices remains challenging. This paper introduces a novel lightweight model for heart sound classification based on broadcast residual learning, which can be seamlessly deployed in portable health monitoring devices for real-time cardiac auscultation. Comparative experiments validate the model’s efficacy, achieving 89.1% accuracy and 89.7% F1 score with just 8.05 K parameters and 10.6 M MACC, showcasing superior performance within constrained complexity.