A Federated Learning Paradigm for Heart Sound Classification.

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Data Privacy and Security

Background:

• Cardiovascular diseases (CVDs) are a leading cause of mortality, necessitating early and accessible diagnostic methods.
• Automated heart sound auscultation offers a promising, non-invasive approach for continuous cardiac monitoring.
• Existing methods face challenges in balancing the need for large datasets with stringent user data privacy requirements.

Purpose of the Study:

• To introduce and evaluate a federated learning (FL) framework for heart sound classification, prioritizing user data privacy.
• To investigate the impact of data distribution across institutions on FL model performance in cardiac diagnostics.
• To demonstrate the feasibility and effectiveness of FL for automated heart sound analysis.

Main Methods:

• Development and implementation of a federated learning (FL) framework tailored for heart sound classification.
• Conducting experiments with real-world heart sound data from multiple collaborative institutions.
• Analysis of model performance under varying non-identically and independently distributed (Non-IID) data scenarios.

Main Results:

• The proposed FL framework effectively classifies heart sounds while preserving user privacy.
• Model quality and learning patterns were analyzed, showing the influence of data distribution.
• A strategy of globally sharing data was found to significantly improve performance in Non-IID settings.

Conclusions:

• Federated learning (FL) represents a novel and effective approach for privacy-preserving automated heart sound analysis.
• The study validates the feasibility of FL in real-world medical applications for CVD diagnosis.
• Global data sharing strategies enhance FL model robustness and accuracy, particularly with heterogeneous data.