Intelligent Recognition of Anomalous Behaviors in Medical Insurance Through Deep Learning

Abstract

Medical insurance fraud represents a critical financial burden on healthcare systems globally, with annual losses exceeding billions of dollars. This paper presents a comprehensive investigation into deep learning-based anomaly detection frameworks designed to identify fraudulent behaviors in medical insurance claims. The proposed approach employs a hybrid architecture where: (1) deep factorization machines generate embedding-based features from sparse categorical data, (2) heterogeneous graph neural networks extract relational features from provider-patient-pharmacy networks, and (3) these complementary feature representations are integrated through a two-stage ensemble framework combining cost-sensitive XGBoost, weighted stacking, and focal loss optimization. The final detection pipeline consists of parallel feature extraction modules (DFM embeddings, GNN node representations, temporal CNN encodings) feeding into a meta-ensemble that produces anomaly scores. The framework incorporates explainable AI mechanisms through SHAP and attention-based interpretability, enabling transparent decision-making for regulatory compliance. Extensive experimental validation demonstrates superior performance in detecting complex fraud patterns across multiple dimensions including visit frequencies, billing amounts, and prescription combinations. The adaptive learning mechanisms enable continuous model evolution to address emerging fraud typologies while maintaining interpretability for audit personnel.