FedProx-Based Federated Transfer Learning for Efficient Model Personalization in Healthcare

Federated Learning (FL) facilitates collaborative model training among multiple clients at diverse locations while preserving data privacy, making it particularly suitable for healthcare applications. Nevertheless, conventional (FL) techniques often face challenges such as significant communication overhead and unstable convergence, especially when dealing with heterogeneous data. To address these challenges, we propose a FedProx-based Federated Transfer Learning (FTL) framework that leverages pre-trained models to improve convergence stability with less training. The proposed method introduces a proximal regularization term that penalizes large updates to local models, thereby ensuring model robustness across different hospitals. The use of pre-trained models further reduces the amount of local data needed for training, enabling effective personalization with limited datasets. Experiments on a synthetic healthcare dataset demonstrate that our method achieves comparable performance to traditional FL methods while significantly improving convergence and learning capability, making it suitable for resource-constrained healthcare environments.