Abstract

This paper proposes a unified and lightweight framework for multi-class cardiovascular disease detection using heart sound. It is designed to align signal preprocessing and model
architecture within a coherent pipeline. Heart sound recordings are standardised through simple signal conditioning and transformed into fixed-size log-mel spectrograms, providing a consistent time–frequency representation for learning. On this basis, a compact multi-branch neural network, RDLINet+, is introduced, which enhances a strong lightweight baseline through channel-wise feature refinement. The framework is evaluated on a balanced five-class heart sound dataset under a strict and consistent protocol, alongside CNN, RNN, and baseline RDLINet models. Results show that the proposed approach achieves superior accuracy, improved inter-fold stability, and clearer class discrimination, particularly for clinically similar valvular conditions. The findings highlight that effective coordination between representation and model design, rather than increased complexity, can yield robust and high-performing heart sound classification, supporting practical computer-aided auscultation.