Cloud-Based Federated Learning Implementation Across Medical Centers.

Area of Science:

• Health Informatics
• Machine Learning
• Data Science

Background:

• Machine learning (ML) in healthcare is hindered by data-sharing concerns, limiting the size of training datasets.
• Federated learning (FL) offers a solution by enabling model training across multiple institutions without direct data sharing.

Purpose of the Study:

• To explore and evaluate various federated learning implementations in a healthcare context.
• To assess the effectiveness of FL using electronic health record (EHR) data from two academic medical centers.

Main Methods:

• Implemented FL using artificial neural networks (ANNs) and logistic regression (LR) models.
• Tested incremental and cyclic FL models in both simulated and real-world environments using EHR data on a cloud platform.
• Exchanged ML models between institutions via a GitHub repository.

Main Results:

• Cyclically trained ANNs demonstrated a statistically significant 3% performance increase (P < .05).
• Single weight neural network models showed performance improvements in some instances.
• Logistic regression models did not exhibit significant improvements after FL.
• The order of institutional training influenced the overall performance gains.

Conclusions:

• Federated learning can be effectively implemented beyond simulations in healthcare settings.
• Specific FL models, particularly cyclic ANNs, achieved statistically significant performance improvements.
• Further research is required to optimize FL for biomedical applications while ensuring data security and privacy.