Privacy-Preserving Federated Learning for Prognostic Modeling in Rare Diseases: A Scalable Case Study on Kawasaki Disease

Abstract

Predictive modeling in rare diseases faces major challenges, including data scarcity, class imbalance, and strict privacy regulations that limit cross-border collaboration. These challenges are particularly critical in Kawasaki disease (KD) — a rare vasculitis in children — where 10% to 20% of patients are resistant to intravenous immunoglobulin (IVIG), the standard first-line treatment. This significantly increases the risk of coronary artery abnormalities, making early and accurate prediction of resistance to IVIG essential for improving patient outcomes. Our work proposes a Federated Learning approach to address the constraints imposed by security and privacy concerns. We investigate Convolutional Neural Networks as the shared model, collaboratively trained across clients. Coupled with strategies to address class imbalance resulting from the rarity of the condition, the federated approach yielded promising results when evaluated against conventional machine learning models. The proposed approach demonstrated strong performance, achieving 94% accuracy, 93% precision, 89% recall, and 91% F1 score. To ensure robustness and generalizability, an independent dataset was also used, where the proposed model excelled similarly. These results highlight the potential of federated learning to overcome data privacy barriers and provide a scalable, secure solution for predictive modeling in rare diseases, supporting its integration into medical prediction workflows.