Federated learning of predictive models from federated Electronic Health Records

Area of Science:

• Machine Learning
• Healthcare Informatics
• Distributed Computing

Background:

• Large-scale machine learning in healthcare faces challenges with centralized data storage, including privacy risks and scalability issues.
• Decentralized solutions are crucial for analyzing distributed health data across institutions.
• Existing centralized algorithms are impractical for multi-institutional collaborations due to data privacy and logistical constraints.

Purpose of the Study:

• To develop a general decentralized optimization framework for collaborative machine learning on sensitive health data.
• To predict hospitalizations for cardiac events using a distributed algorithm without explicit raw data exchange.
• To enable multiple data holders to converge on a common predictive model while preserving data privacy.

Main Methods:

• The study focuses on the soft-margin l1-regularized sparse Support Vector Machine (sSVM) classifier.
• An iterative cluster Primal Dual Splitting (cPDS) algorithm was developed for decentralized, large-scale sSVM problem solving.
• This distributed learning scheme facilitates collaboration among data holders, ensuring data privacy.

Main Results:

• The cPDS algorithm demonstrated faster convergence than centralized methods for predicting cardiac event hospitalizations from Electronic Health Records.
• cPDS achieved comparable prediction accuracy (Area Under the ROC Curve) to centralized approaches with reduced communication overhead compared to alternative distributed methods.
• Key predictive features for future hospitalizations were identified, aiding in result interpretation and prevention strategies.

Conclusions:

• The developed decentralized cPDS algorithm offers an efficient and privacy-aware solution for large-scale machine learning in healthcare.
• This approach supports multi-institutional collaborations by enabling model training on distributed data without compromising patient privacy.
• The identified predictive features provide valuable insights for clinical decision-making and preventative care in cardiology.