Multi-Feature Decision Fusion Network for Heart Sound Abnormality Detection and Classification.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Heart sounds offer crucial early indicators of cardiovascular diseases.
• Automatic heart sound analysis is vital for prompt disease detection.
• Existing methods face challenges in adapting to diverse diagnostic tasks.

Purpose of the Study:

• To develop a novel, adaptable end-to-end method for heart sound abnormality detection and classification.
• To enhance the discriminative power and fusion of multi-dimensional heart sound features.
• To improve performance in the absence of cardiac cycle segmentation.

Main Methods:

• Developed a Multi-feature Decision Fusion Network (MDFNet) with Multi-dimensional Feature Extraction (MFE) and Multi-dimensional Decision Fusion (MDF) modules.
• MFE module extracts spatial, temporal, and spatio-temporal features.
• MDF module employs deep supervision, decision fusion, and attention mechanisms for enhanced feature discrimination and information integration.
• Introduced an efficient data augmentation technique.

Main Results:

• Achieved 94.44% accuracy and 86.90% F1-score on binary classification.
• Attained a 99.30% F1-score on a five-classification task.
• Outperformed existing state-of-the-art methods in heart sound analysis.

Conclusions:

• The proposed MDFNet demonstrates high performance and adaptability for cardiovascular disease diagnosis using heart sounds.
• The method shows significant potential for clinical application in early disease detection.
• The novel approach effectively addresses limitations of previous end-to-end methods.