Continual learning framework for a multicenter study with an application to electrocardiogram.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Medical Data Privacy

Background:

• Deep learning models require large datasets, often necessitating multicenter collaborations.
• Data sharing in medical research is restricted due to privacy concerns and legal regulations.
• Existing federated learning requires a central server, posing cost and regulatory challenges.

Purpose of the Study:

• To propose a novel continual learning framework for multicenter studies that does not require a central server.
• To address the challenge of privacy preservation in distributed medical data training.
• To prevent catastrophic forgetting in machine learning models trained on evolving datasets.

Main Methods:

• Developed a continual learning framework with a method selection process for diverse datasets.
• Utilized generative adversarial networks to create synthetic data for prospective evaluation.
• Trained arrhythmia detection models on four independent electrocardiogram datasets.

Main Results:

• The proposed framework achieved stable performance (AUROC 0.897) across all datasets without a central server.
• Performance was comparable to traditional federated learning (AUROC 0.901).
• Effectively prevented catastrophic forgetting of previously learned knowledge.

Conclusions:

• The proposed framework offers a viable, privacy-preserving alternative for multicenter medical studies.
• It demonstrates the potential of server-less continual learning in healthcare AI.
• This approach facilitates robust model training on distributed medical data.