IPFL: Interpretable Federated Learning for Personalized Healthcare
Arthur A. Nijdam, Amir Aminifar · IEEE Access · 2025
Federated Learning (FL) enables decentralized training of neural networks across multiple hospitals or patients while preserving data privacy. However, FL schemes typically assume data is independent and identically distributed (IID) while healthcare data can be highly heterogeneous. To address this, we propose Interpretable Personalized Federated Learning (IPFL), a novel framework that allows patients to selectively collaborate with others based on both validation performance and historical collaboration success. By directly inferring patient similarities from data, IPFL enables personalized model training without requiring assumptions about cluster structures or interpolation with a global model. We validate IPFL on two real-world healthcare tasks: epileptic seizure detection and cardiac arrhythmia detection, and show that it achieves state-of-the-art performance. Moreover, our analysis demonstrates that IPFL naturally leads to interpretable collaboration graphs: patients with similar disease characteristics tend to collaborate more frequently.