Healthcare Fraud Detection using Machine Learning

Healthcare fraud in the United States signifies a considerable illicit financial drain with estimations suggesting annual losses amounting to tens of billions of dollars. Such fraudulent activities encompass a wide array of schemes including but not limited to billing for unrendered services, upcoding to receive higher reimbursements, and engaging in unlawful kickback arrangements. Recognizing the criticality of this issue, this research analyzes the utilization of machine learning techniques for the detection of healthcare fraud. Through an analysis of the dataset which amalgamates inpatient, and outpatient claims data with beneficiary information across 558,211 records spanning various dimensions, this study illustrates the application of several ML models including Random Forest, XGBoost, SVM, Isolation Forest, a Deep Learning Model,and a Stacking Ensemble approach. The models are evaluated based on their accuracy, precision, recall, F1 score, and ROC AUCscore with a particular focus on their applicability to healthcare fraud detection. Among the models evaluated, the Stacking Ensemble Model emerged as particularly efficacious, achieving an accuracy of 92.79% and an exceptional ROC AUC score of 96.95%. Incorporating hyperparameter tuning, this study further enhances interpretability and decision-making through SHAP value analysis, offering deep insights into model predictions and feature importances. Additionally, it introduces an innovative real-time healthcare fraud detection pipeline and an automated model retraining framework, ensuring the system remains effective against evolving fraud tactics by continuously adapting and improving.