Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Biomedical Data Science

Background:

• Deep learning excels at identifying complex patterns for medical biomarkers.
• Training deep learning models requires large, diverse datasets, often unavailable at single institutions.
• Sharing patient data for multi-institutional collaboration faces significant privacy and ownership hurdles.

Purpose of the Study:

• To introduce and evaluate federated learning as a privacy-preserving method for multi-institutional medical data collaboration.
• To assess the model quality and generalizability of federated learning compared to centralized approaches.
• To investigate the impact of data distribution on federated learning model performance.

Main Methods:

• Federated learning was implemented across 10 institutions, training models locally without centralizing data.
• Model performance was evaluated against a centralized training baseline.
• Generalizability was tested on external datasets from non-federated institutions.
• The influence of data distribution heterogeneity on model quality was analyzed.

Main Results:

• Federated learning models achieved 99% of the quality of models trained on centrally pooled data.
• The approach demonstrated strong generalizability on external datasets.
• Increased data access through federated learning outweighed potential errors from the collaborative method.
• Federated learning outperformed other collaborative learning techniques.

Conclusions:

• Federated learning offers a viable, privacy-preserving solution for multi-institutional medical data collaboration.
• This approach can lead to the development of robust deep learning models trained on unprecedentedly large datasets.
• Federated learning is poised to significantly impact the advancement of precision and personalized medicine.