Personalized Federated Learning for Institutional Prediction Model using Electronic Health Records: A Covariate...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Federated learning (FL) enables multi-institutional AI collaboration without data sharing, but struggles with non-identically distributed (Non-IID) data.
• Personalized Federated Learning (PFL) addresses Non-IID data by creating client-specific models, yet lacks interpretability regarding data contribution.
• Interpretable personalization is crucial for AI in medical applications to understand data sample contributions.

Purpose of the Study:

• To propose a novel PFL framework, Federated Adjustment of Covariate (FedCov), for interpretable personalization.
• To enable visualization of client contributions within PFL.
• To improve AI model robustness and performance in collaborative medical settings.

Main Methods:

• FedCov estimates propensity scores to model covariate shift among clients via prior FL.
• It learns a final model by weighting training sample contributions based on estimated propensity scores.
• The framework facilitates covariate adjustment for personalized model learning and contribution visualization.

Main Results:

• FedCov achieved an ROC-AUC of 0.750 in predicting in-hospital mortality across 50 hospitals.
• This performance surpassed conventional FL methods (AUC 0.720-0.735) and neared centralized learning (AUC 0.754).
• The method demonstrated interpretable personalization by visualizing client data contributions.

Conclusions:

• FedCov offers a feasible approach for interpretable personalized federated learning in healthcare.
• This framework enhances AI-driven clinical decision support by enabling personalized models for any medical institution.
• The study highlights the potential of FedCov to improve AI reliability and transparency in medical collaborations.