Fraud detection and explanation in medical claims using GNN architectures

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Medical insurance fraud poses a significant financial burden on healthcare systems.
• Traditional fraud detection methods struggle to capture complex inter-entity relationships in healthcare data.
• Graph Neural Networks (GNNs) offer a promising approach for analyzing relational data.

Purpose of the Study:

• To investigate the efficacy of GNNs for detecting fraud in medical insurance claims.
• To model healthcare entities and their interactions as a heterogeneous graph for fraud analysis.
• To compare the performance of different GNN architectures on real-world claims data.

Main Methods:

• Utilized two state-of-the-art heterogeneous GNN architectures: HINormer and HybridGNN.
• Adapted a homogeneous GNN, RE-GraphSAGE, to handle heterogeneous healthcare data.
• Evaluated models on real-world medical claims datasets of varying sizes (small, medium, large) comprising millions of activities.

Main Results:

• HINormer and RE-GraphSAGE achieved the highest F-scores (84% and 83%) on small datasets.
• RE-GraphSAGE and HINormer demonstrated top performance (84% and 81% F-score) on medium datasets.
• HINormer and RE-GraphSAGE yielded the best results (82% and 79% F-score) on large datasets.

Conclusions:

• GNNs, particularly HINormer and RE-GraphSAGE, are highly effective for medical insurance fraud detection.
• The performance of GNN models varies across different dataset sizes.
• Explainability techniques (GNNExplainer, PGExplainer) provide insights into model decisions and medical significance.