Enhancing Healthcare AI with Cross-Silo Personalized Federated Learning on Naturally Split Heterogeneous Data

The potential of Artificial Intelligence (AI) in health-care is unavoidable. However, its success depends on the availabil-ity of large, high-quality datasets. Because of data heterogeneity across institutions and privacy concerns, traditional centralized Machine Learning (ML) approaches often face difficulties in this field. Federated Learning (FL) allows collaborative model training without requiring the transfer of sensitive patient data from the original institution. Recent research in FL within the healthcare domain has predominantly relied on centralized datasets, which do not represent real-time data heterogeneity and made assumptions by random data splitting to different medical client institutions. Additionally, it may be challenging for a single global model to encompass the diverse characteristics of various healthcare settings accurately. This paper examines the application of Personalized Federated Learning (PFL) in realistic cross-silo healthcare scenarios with federated natural split datasets in different medical client institutions. This paper discusses the experiments conducted on brain segmentation, survival prediction, melanoma classification, and heart disease di-agnosis. Our experiments show that the proposed PFL techniques consistently improve local model performance over standard FL strategies by up to 10% in different medical use cases.