Federated Abnormal Heart Sound Detection with Weak to No Labels

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Cardiovascular diseases are a leading cause of death, necessitating early detection.
• Artificial intelligence (AI) offers noninvasive heart sound analysis for cardiovascular health assessment.
• Medical data "data islands" hinder collaborative AI model development due to privacy concerns.

Purpose of the Study:

• To validate a federated learning (FL) framework using a positive-unlabeled (PU) learning strategy for abnormal heart sound detection.
• To address challenges of data scarcity and labeling in cardiovascular health AI.
• To explore vertical-FL for cross-institutional collaboration with heterogeneous data.

Main Methods:

• Implemented a federated learning (FL) framework incorporating a naive positive-unlabeled (PU) learning strategy.
• Utilized vertical-FL to enable collaboration across institutions with diverse medical record features.
• Conducted feature importance analysis using six methods to identify key indicators of abnormal heart sounds.

Main Results:

• Achieved an accuracy of 84% in detecting abnormal heart sounds, comparable to supervised learning methods.
• Demonstrated the efficacy of semisupervised FL in learning from limited positive and abundant unlabeled data.
• Provided practical recommendations for abnormal heart sound detection based on feature importance analysis.

Conclusions:

• The validated FL framework with PU learning offers a robust solution for abnormal heart sound detection.
• This approach effectively mitigates data sharing limitations and reduces the burden of manual data labeling.
• The study advances the application of federated learning in noninvasive cardiovascular health assessment.