HS-EfftNet: Heart Sound Classification Using EfficientNet Integrating Dynamic Convolution and Coordinate Attention

Heart sounds are important physiological signals that reflect the mechanical activity of the heart and are widely used in clinical practice to assist in the diagnosis of various heart diseases. Automatic classification based on heart sounds is of great significance for early screening and remote intelligent diagnosis of heart disease. However, since the characteristics of normal heart sounds are very similar to those of abnormal heart sounds, and heart sound samples are often accompanied by noise interference, existing heart sound classification models still face great challenges in distinguishing different types of heart sounds. To address these issues, this paper proposes a heart sound classification model named HS-EfftNet, based on a log-Mel spectrograms. We introduce dynamic convolution into the classical image classification model EfficientNet to enhance its ability to model complex patterns. Additionally, by incorporating a coordinate attention mechanism, the model is guided to focus on non-noise regions and reduce the impact of noise interference. Experimental results and analysis show that the classification performance of HS-EfftNet is significantly improved compared to the original model, and it outperforms existing mainstream classification models across various evaluation metrics, which provide new insights and methods for further research in this field.