FedPPH: A Federated Personalized and Privacy-Preserving Health Model for Multi-Campus Medical Diagnosis

Abstract

This study addresses the issue that multi-hospital medical data cannot be efficiently shared or jointly modeled under privacy constraints, thereby achieving a higher level of intelligent diagnosis and cross-regional collaborative services. The study proposes the Federated Personalized and Privacy-preserving Health (FedPPH) model. FedPPH introduces an adaptive gradient adjustment mechanism into global optimization. It implements noise perturbation and secure aggregation through a two-layer security strategy composed of differential privacy and homomorphic encryption. This ensures privacy constraints while enabling dynamic optimization updates of each client and improving the stability of overall convergence. The study conducts systematic verification based on four public medical datasets: Data 1, 2, 3, and 4 (Healthcare, Hospital Patient Records, Disease Diagnosis, and Medical Data and Hospital Readmissions). The verification methods include horizontal and vertical verification. In horizontal verification, the same dataset is divided into multiple subsets to simulate different campus environments, examining the model's collaborative effect under homogeneous data conditions. In vertical verification, multi-source datasets with significant structural differences test models' generalization and robustness in heterogeneous scenarios. Research results show that in horizontal validation, FedPPH controls the global variance within the range of 0.019–0.035, significantly outperforming the single-node model's range of 0.038–0.077. In vertical validation, taking the Data2&Data3 combination as an example, the client discrepancy decreases from 0.192 to 0.103, demonstrating stronger consistency. Regarding stability, FedPPH maintains an accuracy of 0.892 under perturbed conditions, which remains high compared to the 0.935 accuracy without perturbation. This study aims to provide medical institutions with a collaborative modeling solution that can improve diagnostic accuracy and ensure data security, laying a foundation for the practical implementation of multi-campus smart healthcare.