A Comparative Analysis of Federated Learning Algorithms for Healthcare: Strengths, Limitations, and Practical Recommendations

Abstract

Federated learning (FL) offers a promising framework for collaboratively training machine learning models across distributed healthcare institutions without exposing sensitive patient data. This decentralized approach aligns well with privacy regulations such as HIPAA and GDPR and addresses challenges posed by siloed electronic health records (EHRs), medical imaging archives, and genomic data. Despite its growing relevance, FL remains difficult to deploy effectively in healthcare due to data heterogeneity, resource imbalance, and privacy constraints. This paper presents a comprehensive comparative analysis of five key FL algorithms—FedAvg, FedProx, SCAFFOLD, FedBN, and pFedMe—evaluated across structured data, imaging, and segmentation tasks using MIMIC-III, MedMNIST, and BraTS datasets. We assess these algorithms in terms of predictive performance, convergence speed, fairness, communication efficiency, and robustness under non-IID settings. Experiments simulate realistic healthcare conditions, including client dropouts, bandwidth constraints, and system variability. Our findings highlight that no single algorithm universally dominates across all tasks. Personalized and variance-reducing methods tend to perform better in diverse data settings, though at the cost of increased complexity. We discuss key trade-offs, scalability limitations, and ethical considerations—including alignment with emerging regulatory frameworks such as the EU AI Act and OECD AI Principles. This paper provides practical guidance for selecting appropriate FL strategies in privacy-sensitive healthcare applications.

Ähnliche Arbeiten