Aggregation Algorithm Selection for Non-IID Federated Learning in Healthcare: Analysis of Convergence, Privacy, and Communication Tradeoffs

Abstract

With Federated Learning (FL), hospitals are able to train models together without moving patient data, which is important due to security laws such as HIPAA and GDPR that make most deployments legally prohibit training models with patient data in a central location. The larger issue is statistical. Hospitals likely have differing patient populations, use differing hardware and imaging methods, and have different disease prevalence, all of which impact how data is distributed. FL degrades in these scenarios. This paper seeks to evaluate six aggregation algorithms FedAvg, FedProx, SCAFFOLD, FedMA, FedBN, and Per-FedAvg and studies their behavior on non-IID clinical data. We assess the convergence rate, communication cost, accuracy under distribution shift, and the algorithms’ privacy mechanisms. SCAFFOLD has the fastest convergence in high heterogeneity (Dirichlet α=0.5), though it has double per round communication overhead, so it is not feasible under limited bandwidth in clinical sites. FedProx, like FedAvg, has a moderate convergence cost and has more stable convergence and is thus a more deployable option. Three privacy-enhancing technologies—DP, secure aggregation, and HE—are analyzed across medical imaging, EHR prediction, and genomic tasks, with accuracy–privacy tradeoffs quantified. A four-dimensional taxonomy organizes the design space. Six open problems are identified: federated foundation models, continual learning, unlearning for GDPR compliance, multimodal fusion, reinforcement learning across trial sites, and FL on edge devices.