Agent-guided pruning and quantization strategy for lightweight heart sound classification

Heart sounds contain a wealth of physiological information about the heart. Automatic heart sound diagnosis plays an important role in the early detection of cardiovascular diseases and assisting clinicians in auscultation. This study developed a lightweight heart-sound classification method using agent-guided pruning and quantization strategy. Specifically, this study designed a hybrid lightweight adaptive mamba (HLAmamba) module, which mainly consists of a smooth-tanh state space model (SSSM) and an adaptive wavelet activation function. SSSM efficiently models long sequences and deeply extracts heart sound signature. The adaptive wavelet activation function (AWAF) highlights the characteristics of heart sound diseases. Combining the two can accurately identify the abnormal heart sounds. The designed lightweight network, the HLAmamba Network (HLAMNet), reduces the number of training parameters and storage. Finally, a pruning and quantization multi-agent was developed to infer the pruning rate and quantization bit number, further compressing the model. The developed method achieved an accuracy of 99.22% on the five-category classification task and 100% on the two-category task. It outperforms other advanced models and has excellent stability and robustness, and has good clinical application prospects.