Personalized Federated Multi-Center Medical Data Analysis with Local and Global Uncertainty

Multi-center brain disease diagnosis is essential for developing effective treatments and improving patient outcomes. However, concerns about data leakage have hindered its development, and data heterogeneity caused by differences in data collection devices and regional population characteristics is a significant challenge. In this paper, we propose a personalized federated multi-center brain disease diagnosis framework based on local and global uncertainty (PFLGU) and evaluate its performance in diagnosing autism and schizophrenia across multiple centers. First, we design two parallel classifiers for each center to generate uncertainty of samples. Second, based on the uncertainty, we utilize the min-max optimization approach for local training. Then, we use the average uncertainty of samples at each center as its weight for federated aggregation. Finally, we obtain a personalized model at each center that effectively incorporates information from other centers. Experimental results demonstrate that the proposed PFLGU framework improves multi-center diagnostic performance for autism and schizophrenia.