International Perspectives on Rural Medical Insurance Fraud Detection: A Machine Learning Approach with SHAP Interpretability

This paper presents a machine learning-based framework for detecting rural medical insurance fraud, utilizing multi-source data such as beneficiary information (KYC), hospitalization claims, outpatient claims, and provider fraud labels. The study integrates, processes, and engineers features from these diverse data sources to train multiple machine learning models, including LightGBM, random forest, and gradient boosting. Comparative experimental results indicate that the LightGBM model outperforms other models across several evaluation metrics, including accuracy, precision, recall, F1-score, and AUC. Additionally, SHAP interpretability analysis identifies the total amount of insurance claims, the mean duration of claims, and the mean number of diagnoses as critical factors influencing fraud detection. This research provides a robust technical solution for rural medical insurance fraud detection, offering significant potential for application in various international healthcare systems, ensuring the protection of insurance funds on a global scale.