Privacy-Preserving Federated Graph Neural Network for Collaborative Health Insurance Fraud Detection with Class-Imbalance Mitigation

Abstract

Healthcare insurance fraud is a significant worldwide cause of financial loss for spenders, providers, and beneficiaries, and has the potential to negatively impact the sustainability of healthcare financing systems. Present automated fraud-detection programs often only detect fraud within a single insurer's database, and regulatory and competitive mandates preclude insurers from sharing health-sensitive patient-level data directly with each other, since this can be used to identify fraud. This paper presents FedPPG-Net, a Privacy-Preserving Federated Graph Neural Network framework, which allows multiple Healthcare Insurance Companies (HICs), hospitals and third-party administrators (TPAs) to jointly train a shared fraud-detection model without sharing raw claims data. All participating institutions build a local heterogeneous graph using claims, providers, physicians, and beneficiaries, and train a Graph Attention Network/GraphSAGE encoder with a graph-aware minority-oversampling module (GraphSMOTE) and the focal-loss objective to handle the highly imbalanced nature of fraud labels. Under the differential-privacy framework (DP-SGD) local model updates are secured by using gradient clipping and Gaussian-noise injection per example, before being combined at a central server using secure aggregation and FedAvg/FedProx. The framework is tested on the publicly available Kaggle Healthcare Provider Fraud Detection Analysis data set, obtained from U.S. CMS Medicare claims data, split into non-IID client silos, mimicking a realistic multi-insurer deployment. Experimental results demonstrate that the proposed model achieves an F1 score of 0.803, an AUC-ROC of 0.918 and an AUC-PR of 0.779, outperforming the centralized and federated baselines, while also having a bounded utility loss under strong (ε ≤ 4) privacy guarantees. Both the class-imbalance mitigation module and the graph-based relational encoder are shown to play important roles in the detection performance through ablation studies, and a privacy-utility analysis provides practical insights on implementing the framework in real-life multi-institutional healthcare environments.