Introduction and Comparison of Novel Decentral Learning Schemes with Multiple Data Pools for Privacy-Preserving ECG...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Artificial intelligence (AI) and machine learning (ML) offer significant innovations across various fields.
• Medical applications of AI/ML face challenges due to data privacy concerns and stringent legal regulations.
• Decentralized knowledge-based methods can mitigate privacy issues by avoiding data centralization.

Purpose of the Study:

• To apply and compare 6 decentralized machine learning algorithms for 12-lead ECG classification against conventional centralized methods.
• To evaluate the trade-offs between classification performance and privacy preservation in medical AI.
• To identify optimal decentralized algorithms for privacy-preserving AI in healthcare.

Main Methods:

• Implementation of 6 distinct decentralized machine learning algorithms.
• Application of these algorithms to 12-lead ECG datasets for classification tasks.
• Comparison of decentralized methods against a standard centralized machine learning model.

Main Results:

• Federated learning (FL) demonstrated a minor decrease in classification performance (AUROC -0.054) compared to centralized models, while significantly enhancing privacy.
• A weighted FL variant (AUROC -0.049) and an ensemble method (AUROC -0.035) surpassed standard FL performance.
• A novel batch-wise sequential learning scheme yielded the best performance (AUROC -0.036) relative to the baseline.

Conclusions:

• Decentralized machine learning, particularly federated learning, presents a viable approach for privacy-preserving AI in medicine.
• Advanced FL techniques, including weighted variants and ensemble methods, improve upon standard FL performance.
• The batch-wise sequential learning scheme shows promise for optimal performance in privacy-preserving ECG classification.