Learning Across the Divide: Personalised Federated Learning for Robust Clinical Modelling Under Data-View...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Health Informatics

Background:

• Federated Learning (FL) facilitates collaborative clinical modeling using distributed electronic health records (EHRs) while preserving patient privacy.
• Data-view heterogeneity, arising from variations in medical practices and documentation, presents a significant challenge to standard FL methods.
• Existing solutions often involve complex, lossy data preprocessing and manual harmonization, limiting scalability and personalization.

Purpose of the Study:

• To introduce the Personalised Attention-based Federated Graph Network (PAFNet), a novel FL framework designed to address data-view heterogeneity in clinical settings.
• To enable effective and scalable collaborative model training across institutions with diverse data feature sets without extensive manual preprocessing.

Main Methods:

• PAFNet maps heterogeneous client data-views into a shared latent space using client-specific projection layers.
• A personalized adaptation mechanism with trainable parameter masks allows clients to selectively use relevant global model parameters.
• This approach preserves local data specificity and avoids the need for extensive data harmonization.

Main Results:

• PAFNet demonstrated superior performance compared to state-of-the-art FL methods on heterogeneous datasets (CURIAL, eICU, MIMIC-III).
• The framework showed strong generalization capabilities even with significant differences in client feature sets.
• PAFNet effectively enabled personalization and cross-institutional knowledge sharing.

Conclusions:

• PAFNet offers a robust and scalable solution for federated clinical model training in environments with data-view heterogeneity.
• The proposed method overcomes limitations of existing approaches by enabling effective personalization and reducing reliance on manual data harmonization.