A novel 1-D densely connected feature selection convolutional neural network for heart sounds classification

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Heart sound auscultation is a primary, non-invasive method for cardiovascular disease detection.
• Existing methods require continuous improvement for accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework, Dense-FSNet, for accurate identification of abnormal heart sounds.
• To enhance cardiovascular disease detection through advanced signal processing techniques.

Main Methods:

• A novel Dense Feature Selection Convolution Network (Dense-FSNet) framework was developed, incorporating circular dense connectivity modules (Clique Blocks).
• Squeeze-and-Excitation blocks and a multi-scale attention mechanism were integrated to optimize feature representation and parameter efficiency.
• Data balance technology was employed to enhance model robustness.

Main Results:

• The Dense-FSNet framework achieved adaptive spatial feature fusion by integrating high-level and low-level features.
• The model demonstrated notable performance improvements through the implemented attention mechanisms and data balancing.
• Performance was assessed using accuracy, specificity, sensitivity, and AUC values.

Conclusions:

• The proposed Dense-FSNet model achieves state-of-the-art results in classifying and detecting abnormal heart sounds.
• This framework shows significant potential for improving cardiovascular disease diagnosis using heart sound analysis.