Abstract
Healthcare fraud, particularly organized gang activity, poses a significant challenge to insurance systems, costing billions annually. Traditional detection methods often fail to capture the complex, multi-scale patterns of conspiracy. We present RDPGL (Risk Diffusion-based Parallel Graph Learning), a novel framework that integrates local entity interactions with global claim-level risk diffusion. By modeling claims as nodes in a relation graph and propagating risk scores, RDPGL identifies sophisticated fraud patterns that local-only models miss. Evaluated on a large-scale real-world dataset, RDPGL demonstrates superior performance, achieving over 95% precision for top-tier suspicious claims and outperforming state-of-the-art graph neural network baselines.
Introduction
Healthcare fraud is a staggering $300 billion annual problem in the U.S. alone. While early detection systems focused on "lone wolf" fraudsters, modern criminals operate like sophisticated enterprises—organized gangs that move quickly and use deceptive, ever-changing tactics to hide their tracks.
Existing Graph Neural Networks (GNNs) often struggle with this because they typically focus on local context (e.g., a single suspicious relationship between one patient and one hospital). They miss the global context: the "conspiracy patterns" where multiple claims across different entities are linked by a shared, hidden logic. The challenge is twofold: fraudsters are far-reaching and mobile, and their behavioral patterns evolve rapidly, making static rules obsolete.
Methodology: The RDPGL Framework
The proposed RDPGL framework employs a "divide and conquer" strategy via two parallel branches to capture both granular interactions and broad network patterns.
Local Entity Graph Learning
This branch utilizes a heterogeneous graph attention network to model relationships within a beneficiary-provider graph. It learns "local" features—how specific patients and doctors interact. An attention mechanism is used to weigh which neighbors and specific claims are most relevant to a node's representation, allowing the model to focus on suspicious local clusters.
Global Claim Graph Learning
The core innovation of RDPGL is the construction of a claim-claim relation graph, where nodes represent individual insurance claims. An edge exists between two claims if they share the same doctor or patient.
- Community-aware Risk Diffusion: To capture the "contagion" of fraud, a modified PageRank-style algorithm spreads "risk scores" from known fraudulent claims to their neighbors based on attribute similarity and "hop distance."
- Pre-training: A vanilla Graph Convolutional Network (GCN) is used to pre-train these global risk representations before the final task.
Latent Feature Aggregation
The local and global embeddings are concatenated and processed through a Multi-Layer Perceptron (MLP) to calculate the final fraud probability. This dual-view approach allows the model to simultaneously analyze individual entity behavior and overarching criminal network structures.
Experimental Evaluation
The model was evaluated on a real-world medical insurance dataset from the National Healthcare Security Administration (NHSA), containing over 0.55 million claims and 0.2 million beneficiaries. The data includes expert-annotated labels for both individual and organized "gang-level" fraud across inpatient and outpatient scenarios.
Performance and Precision
RDPGL consistently outperformed seven state-of-the-art baselines, including GAT, DCI, and BWGNN. In outpatient scenarios, RDPGL achieved an AUC of 0.8713 and an F1-score of 0.7982. Notably, for the top 1% of predicted fraud claims, the model achieved over 95% precision, which is critical for real-world investigative efficiency.
Ablation Study: The Necessity of Parallelism
Ablation studies confirmed that both local and global contexts are vital for detecting organized fraud. Removing either branch led to a significant drop in performance across all data subsets (Inpatient-Outpatient, Inpatient-Inpatient, etc.).
Conclusion and Future Work
RDPGL effectively addresses the limitations of traditional GNNs in healthcare fraud detection by treating fraud risk as a diffusive property across a global claim graph. While the model shows high robustness, future work will focus on addressing data imbalance in inpatient records and explicitly modeling the temporal evolution of fraud gangs to better track how these criminal networks form and dissolve over time.
Summary updated