Comparative Study of Federated Learning Techniques for Healthcare Applications

Abstract

The exponential growth of healthcare data and the concurrent need for privacy-preserving machine learning models have propelled Federated Learning (FL) into the forefront of healthcare artificial intelligence research. FL enables multiple medical institutions to collaboratively train AI models without centralizing sensitive patient data. This manuscript presents a comparative study of popular federated learning techniques, including FedAvg, FedProx, FedSGD, and Scaffold, evaluating their effectiveness in healthcare scenarios such as disease diagnosis, patient monitoring, and medical image classification.

Using publicly available datasets including MIMIC-III and COVIDx, we simulate multi-institution federated settings and perform detailed statistical analysis on metrics such as accuracy, convergence time, communication cost, and privacy leakage risk. The results show that while FedAvg is communication-efficient and robust, Scaffold offers superior convergence in heterogeneous data environments. FedProx is particularly useful under non-IID conditions prevalent in clinical data. This study highlights the trade-offs between algorithmic complexity, performance, and privacy guarantees, concluding with suggestions for FL technique selection based on specific healthcare use cases.