JASSNet: Heart and lung sound separation network based on joint attention mechanism and Semi-Supervised learning

High-quality heart and lung sounds are essential for the automated diagnosis of cardiovascular or respiratory diseases. However, chest sounds recorded by stethoscopes often contain a mixture of heart and lung sounds, and the clinical measurement of high-quality cardiopulmonary sounds is costly, with labeled data being scarce. To address these challenges, this study proposes a heart-lung sound separation network based on a joint attention mechanism and semi-supervised learning (JASSNet). Specifically, a convolutional module is designed to extract fine-grained features, combined with both global and local attention mechanisms to enhance information interaction between features. A gating mechanism is employed to extract key heart and lung sound features. Finally, clinical validation is performed by comparing different semi-supervised learning strategies. The results demonstrate that the proposed model performs excellently on both simulated and clinical datasets. Notably, in clinical trials, the accuracy of heart rate (HR) and respiratory rate (RR) improved by over 10% and 15%, respectively. This study not only shows potential in the preprocessing steps of health monitoring systems but also holds significant clinical application prospects.