Personalized Federated Graph Learning on Non-IID Electronic Health Records.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain valuable latent disease patterns for proactive healthcare.
• Federated graph learning extracts patterns from distributed EHRs without raw data sharing.
• Non-independent and identically distributed (Non-IID) EHR data presents challenges like imbalance, reducing global model effectiveness.

Purpose of the Study:

• To introduce PEARL, a personalized federated learning framework for disease prediction on Non-IID EHRs.
• To address data imbalance and enhance healthcare decision-making accuracy.
• To ensure privacy during federated learning model updates.

Main Methods:

• PEARL utilizes disease diagnostic code and admission record attention for patient embedding extraction.
• Integrates self-supervised learning within a federated framework for hierarchical disease prediction.
• Employs a fine-tuning scheme for personalized client models and differential privacy for secure updates.

Main Results:

• Extensive experiments on the MIMIC-III dataset demonstrate PEARL's effectiveness.
• PEARL achieves competitive results compared to existing baseline methods.
• The framework successfully handles Non-IID EHR data for improved disease prediction.

Conclusions:

• PEARL offers a robust solution for disease prediction in decentralized, Non-IID EHR settings.
• The personalized federated learning approach enhances model performance and privacy.
• This framework advances the application of AI in healthcare for more precise decision-making.