Enabling end-to-end secure federated learning in biomedical research on heterogeneous computing environments with...

Area of Science:

• Biomedical Machine Learning
• Federated Learning
• Data Privacy

Background:

• Large-scale biomedical machine learning (ML) projects require secure collaboration across institutions.
• Existing federated learning (FL) environments face challenges in ensuring data confidentiality, particularly with protected health information.
• Cross-institutional research is hindered by administrative and data security boundaries.

Purpose of the Study:

• Introduce APPFLx, a low-code, user-friendly FL framework designed for secure, cross-institutional biomedical ML.
• Enable secure end-to-end communication, privacy-preserving functionality, and robust identity management.
• Facilitate the deployment of FL into existing computational infrastructures without modification.

Main Methods:

• Developed APPFLx, a framework supporting secure FL experiment setup and execution.
• Demonstrated APPFLx utility through two biomedical case studies: age prediction from ECG and COVID-19 detection from radiographs.
• Utilized heterogeneous computing resources, including on-premise and cloud facilities, for secure model training.

Main Results:

• APPFLx successfully facilitated secure, cross-institutional training of ML models on sensitive biomedical data.
• Federated learning models trained using APPFLx demonstrated enhanced generalizability and performance compared to centralized models.
• Data remained protected throughout the training process, ensuring patient privacy.

Conclusions:

• APPFLx is an effective and easy-to-use framework for accelerating biomedical research across organizations.
• The framework enhances collaboration on large datasets while maintaining robust protection of private medical data.
• APPFLx overcomes administrative barriers, enabling secure and efficient federated learning in healthcare settings.