Investigating the impact of feature engineering and machine learning model selection for real-world fraud detection systems in healthcare insurance claims

The objective of this study is to create and assess an extensive machine learning framework for identifying healthcare fraud in the National Health Insurance Scheme (NHIS) claims, targeting the significant financial losses and degradation of patient care resulting from fraudulent practices. This work examined 20,388 NHIS medical claim data exhibiting phantom billing, incorrect diagnoses, and ghost enrollee fraud trends. A systematic feature engineering approach increased 8 initial characteristics to 27 engineered features, encompassing temporal patterns, financial abnormalities, medical classifications, and indicators of patient behavior. Six machine learning algorithms were assessed: Random Forest, Logistic Regression, Gradient Boosting, XGBoost, Support Vector Machine, and Neural Network, utilizing extensive performance criteria such as accuracy, AUC, calibration quality, and demographic fairness analysis. Gradient Boosting attained the highest test AUC of 0.9213 with an accuracy of 80.11%, whilst XGBoost exhibited superior computational efficiency (0.71 seconds training time) alongside competitive performance (AUC: 0.9187, accuracy: 80.48%). Financial variables predominantly influenced fraud detection judgments, with daily billing rates (AMOUNT_PER_DAY: 0.55) and total billed amounts (0.36) contributing to 91% of model predictions. Significant calibration difficulties were detected across models, with minor demographic bias noted. Ensemble tree-based algorithms routinely surpass alternative approaches in the identification of healthcare fraud. Nevertheless, the primary dependence on financial attributes can cause vulnerabilities to sophisticated fraud schemes that keep accurate billing amounts while capitalizing on weaknesses in medical coding. This research offers healthcare administrators actionable insights for the implementation of real-time fraud detection systems, emphasizing the necessity of balancing detection accuracy with computational efficiency and the enhancement of medical coding analysis capabilities.