Federated Learning with Adaptive Intermediate Model Selection for Predicting IVIG Resistance in Kawasaki Disease

Kawasaki disease (KD), a rare pediatric illness affecting children under five, is treated with intravenous immunoglobulin (IVIG). But 10–20% of patients are resistant to IVIG, and these resistant kids face a higher risk of coronary artery abnormalities. Identifying resistance early is vital, yet data scarcity, class imbalance, and the disease’s rarity necessitate nationwide collaboration, which is often hindered by country-specific privacy policies. Federated learning (FL) provides a practical way for different parties to collaborate on training a model while keeping their raw data private and secure. To enhance model adaptability across diverse clinical populations, we propose an adaptive intermediate model selection strategy in federated learning. Each client retains the version—global or locally fine-tuned—that performs best on its own data, using customizable performance metrics such as F1-score or recall. The system was implemented using the Flower FL framework, with three simulated clients and a shared convolutional neural network (CNN) architecture. Experiments demonstrated that the global model achieved stronger performance than conventional models, and several clients obtained further gains by selecting intermediate models aligned with their data. This approach introduces a novel balance between worldwide collaboration and local personalization in FL, offering a flexible and clinically meaningful solution for IVIG resistance prediction.