Machine Learning for Real-Time Anomaly Detection of Phantom Billing in Health Insurance Claims: An Isolation Forest Approach with Empirical Validation

Abstract

Phantom billing remains one of the most pervasive forms of healthcare fraud, resulting in substantial financial losses for health insurance providers and undermining the integrity of healthcare reimbursement systems. Conventional fraud detection approaches predominantly rely on rule-based auditing and manual claim reviews, which are time-consuming, difficult to scale, and ineffective against emerging fraudulent patterns. The growing volume and complexity of electronic health insurance claims has created a need for intelligent, real-time anomaly detection techniques. This study proposes a machine learning framework for real-time anomaly detection of phantom billing using the Isolation Forest algorithm. The framework integrates claims data preprocessing, feature engineering, anomaly scoring, and empirical validation to identify abnormal billing behaviors without requiring pre-labeled fraud datasets. Relevant claim attributes, including billing frequency, claim amount, diagnosis and procedure codes, provider utilization patterns, and patient service history, are analyzed to distinguish legitimate claims from potential fraud. The model is evaluated using precision, recall, F1-score, ROC-AUC, and PR-AUC, and compared against four baseline anomaly detection methods. The proposed framework provides a scalable, computationally efficient approach for strengthening healthcare claims verification and reducing fraudulent reimbursements. The findings demonstrate the potential of unsupervised machine learning to enhance healthcare fraud detection while establishing a practical foundation for future explainable AI and real-time fraud prevention systems.