Heart sound signal annotation method based on end-to-end and channel attention mechanism

Abstract Heart sound signals reflect cardiovascular function, making their classification essential for detecting, treating, and preventing cardiovascular diseases. Heart sound segmentation, a crucial step in classification, involves identifying the first and second heart sounds, as well as systolic and diastolic signals. Existing segmentation methods primarily rely on envelope features; however, they fail to capture the intrinsic temporal characteristics of the cardiac cycle, leading to suboptimal utilization of heart cycle duration. This study integrates a convolutional neural network with squeeze-and-excitation network modules to extract spatial features and employs a bidirectional long short-term memory network to analyze temporal patterns in heart sound signals. We frame heart sound segmentation as a signal annotation task, enabling end-to-end annotation. The experimental results show that the accuracy of the proposed method is 96.23%, a sensitivity is 96.34%, a accuracy is 96.47%, a specificity is 97.15%, and the F1 score is 97.62% in the heart sound signal dataset. The F1 score is 1.17% higher than that of the comparison method.