A Deep Learning Model for Heart Sound Classification Fusing Time-Frequency Features.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Cardiovascular diseases (CVDs) pose a significant global health risk.
• Accurate phonocardiogram (PCG) signal classification is vital for early CVD diagnosis.
• Existing models often analyze time or frequency domains separately, limiting diagnostic potential.

Purpose of the Study:

• To develop an advanced model for automatic PCG signal classification.
• To overcome limitations of single-domain analysis by integrating time and frequency features.
• To improve classification accuracy and robustness for enhanced CVD diagnosis.

Main Methods:

• Proposed a novel end-to-end dual-branch deep learning architecture.
• Time-domain branch: 1D CNN with Transformer blocks for dynamics and dependencies.
• Frequency-domain branch: ResNet on Mel-spectrograms for spectral patterns.
• Utilized a bidirectional cross-attention fusion module for feature interaction.
• Employed transfer learning for robust performance on diverse datasets.

Main Results:

• Achieved state-of-the-art (SOTA) performance across multiple public datasets.
• Attained 98.86% accuracy and 97.19% F1-score on the 2016 PhysioNet Challenge dataset.
• Significantly outperformed existing baseline methods in heart sound classification.

Conclusions:

• The dual-branch fusion model offers a superior framework for heart sound classification.
• Demonstrates potential for highly accurate automated diagnostic tools for CVDs.
• Supports enhanced early detection and improved clinical outcomes in cardiovascular medicine.