Utilization Analysis and Fraud Detection in Medicare via Machine Learning

Healthcare fraud and overutilization pose significant challenges in the United States, leading to substantial financial losses and compromised patient care. Medicare, a vital federal healthcare program, is particularly susceptible to such abuses. With over 63 million Americans enrolled in Medicare and growing expenses, the need for effective fraud detection is paramount. Traditional methods relying on manual audits have proven insufficient, allowing a significant portion of fraudulent activity to go undetected. Machine learning (ML), however, has gained significant attention in recent years due to its potential for improving the efficiency of fraud detection and prevention. Nonetheless, there are several issues with the existing studies utilizing ML that limit their effectiveness. The most common issue is the heavy reliance on the List of Excluded Individuals and Entities (LEIE) from the Office of the Inspector General for model training and evaluation. Apart from the severe class imbalance issue (with a fraud rate between 0.038% and 0.074%), another notable problem associated with using the LEIE dataset is that many of the providers listed there were prosecuted due to overt and deliberate fraudulent billing. Consequently, using this dataset to train ML models can help detect brazen, outlandish billing patterns, but would be unable to pinpoint instances of more subtle fraud from which a majority of the financial loss and waste occurs. In this paper, we leverage the experience of seasoned physicians and medical billers to create a labeled dataset that overcomes the issues of class imbalance and the exclusive focus on overtly fraudulent providers. We leverage our access to domain knowledge by focusing on the field of ophthalmology. Additionally, using the labeled dataset, we conduct a comparative study of various machine learning models for the task of predicting Medicare overutilization within ophthalmology. The results indicate that our proposed ensemble outperforms individual models such as extreme gradient boosting and multilayer perceptron in detecting overutilization, achieving Area Under the Receiver Operating Characteristic Curve (AUROC score) of 0.907. By deploying the stacking ensemble model, our paper estimates nationwide and jurisdiction-specific overutilization rates, revealing that approximately 8.6% of ophthalmologists engaged in overutilization practices in 2021. We also highlight potential monetary losses of $437.1 million attributed to overutilization activities within ophthalmology for that year alone. Furthermore, feature importance analysis using SHAP (SHapley Additive exPlanations) values provides insights into the key factors influencing the models overutilization predictions. Notably, the ratio of total Medicare payments to the total number of patients emerges as a crucial feature in identifying potential overutilizers.