Enhancing Fraud Detection in Healthcare through Hybrid Machine Learning Ensembles and Optimization

Abstract

Healthcare fraud is a widespread global problem that inflates costs, compromises patient safety, and weakens the integrity of healthcare systems. It involves deceptive practices such as billing for services not rendered, duplicate requests, and misrepresenting patient or service details, causing financial loss. As such, early fraud detection is vital. This paper aims to enhance healthcare fraud detection through hybrid Machine Learning (ML) approaches based on Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT), with and without optimization. We constructed hybrid models by majority vote and applied optimization using Principal Component Analysis (PCA) with hyperparameter tuning (random search). The models were evaluated for precision, accuracy, recall, F1-score, Kappa value, specificity, Mean Absolute Error (MAE), and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) on the Medicare and Medicaid Services (CMS) datasets. The results show that hybrid models generally outperform individual base models, with the RF+DT combination achieving the highest accuracy (0.9597) under standard preprocessing. Optimization methods yielded more balanced results, though with a minor decrease in overall metrics; among these, the LR+RF combination stood out with the highest accuracy (0.9224) and F1-score (0.6161). AUC analysis showed that combining the three classifiers, particularly SVM+RF+DT, achieved the highest discriminative capability (AUC of 0.90 without optimization, 0.85 with optimization). The implications confirm that hybrid ensembles, even with moderate optimization trade-offs, improve predictive robustness and accuracy in healthcare fraud detection.