Heart Sound Abnormality Detection From Multi-Institutional Collaboration: Introducing a Federated Learning Framework.

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Early diagnosis of cardiovascular diseases is critical for effective medical practice.
• Artificial intelligence (AI) offers potential for non-invasive heart sound auscultation but faces challenges with data privacy and scarcity.
• Current AI models often require large datasets, raising privacy concerns and limiting multi-centre collaborations.

Purpose of the Study:

• To propose federated learning (FL) optimization strategies for multi-centre heart sound databases.
• To address privacy issues in AI-driven cardiovascular diagnostics.
• To improve the interpretability and data scarcity challenges in AI models for medical data.

Main Methods:

• Utilized horizontal federated learning (FL) to align feature spaces across institutions without data leakage.
• Employed vertical FL to enhance model interpretability and overcome data scarcity.
• Developed a federated framework for multi-centre heart sound analysis.

Main Results:

• The proposed FL framework achieved strong performance in detecting heart sound abnormalities while ensuring patient privacy.
• Federated feature spaces improved the balance between interpretability and data privacy.
• Demonstrated the effectiveness of FL in multi-centre settings for cardiovascular diagnostics.

Conclusions:

• Federated learning shows significant potential for transitioning AI from research to clinical application in smart medical systems.
• The developed FL strategies effectively address privacy concerns and data limitations in heart sound analysis.
• This work paves the way for broader applications of FL in federated smart healthcare.