Towards federated transfer learning in electrocardiogram signal analysis

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnosis
• Federated Learning in Clinical Settings

Background:

• Current artificial intelligence (AI) models show high performance on healthcare data but often rely on assumptions unsuitable for clinical practice.
• Clinical AI deployment faces challenges with data heterogeneity, distributed sources, and patient privacy concerns.

Purpose of the Study:

• To propose a method for training robust electrocardiogram (ECG) classifiers that handle disparate data dimensions and institutional distribution.
• To develop a privacy-preserving federated learning framework for ECG analysis.
• To introduce a method for creating federated datasets from centralized data.

Main Methods:

• Utilized autoencoders combined with federated learning to address highly heterogeneous modeling problems.
• Employed the Massachusetts Institute of Technology Beth Israel Hospital Arrhythmia dataset, Computing in Cardiology 2017 challenge dataset, and PTB-XL dataset.
• Developed a federated dataset creation method from centralized data.

Main Results:

• Achieved an accuracy of 73.0%, precision of 66.6%, recall of 73.0%, and F1 score of 69.7% with an encoding dimension of 1000.
• Demonstrated the feasibility of federated learning for ECG classification across multiple institutions.
• Highlighted that relaxing common training assumptions complicates the process but yields more realistic performance estimates.

Conclusions:

• The proposed federated learning method effectively handles heterogeneous and distributed ECG data while preserving patient privacy.
• Results suggest that current AI performance estimates in healthcare may be overestimated due to unaddressed clinical data complexities.
• This approach offers a more realistic framework for evaluating and deploying AI in clinical environments.