Personalized Federated Learning via Moreau Envelopes for Collaborative Healthcare Data Analysis

Nowadays, the Federated Learning (FL) has emerged as a transformative approach for collaborative healthcare data analysis by increasing the need to harness large and various datasets while preserving the patient privacy. However, the existing models facing challenges with non-Independent and Identically Distributed (non-IID) settings due to unequal data distributions across patients which led to slower convergence. Hence, this research proposes Personalized Federated Learning via Moreau Envelopes (pFedMe) to allow each client for maintaining a customized model tailored to its local data distribution. Initially, the input images are collected from three datasets namely TB Chest X-ray, Diabetic Retinopathy, and Brain Tumor respectively. Then, the data is fed into SHapley Additive exPlanations (SHAP) to select features by quantifying their individual contribution to a model’s prediction. After that, Synthetic Minority Oversampling Technique (SMOTE) is employed to handle the class imbalance by generating synthetic samples. Then, Residual Network (ResNet) and differential privacy are used to classify the medical images and protect sensitive medical data respectively. Finally, the pFedMe is integrated with adaptive aggregation to handle the real-world complexity of decentralized healthcare data. From the results, the proposed pFedMe achieved better results based on accuracy (98.5%) when compared to existing Microservices-based Distributed Learning (MDL) model.