Fraud Detection with a Limited Number of Known Fraudulent Medicare Providers

Medical claims fraud is a major contributor to increased healthcare costs, but the negative impact can be lessened through effective fraud detection. In this paper, we combine Medicare provider utilization and payment data from 2012 to 2015 with corresponding fraud labels from the List of Excluded Individuals/Entities (LEIE) database. We demonstrate the effectiveness of detecting Medicare fraud with a limited number of known perpetrators, leading to severe class imbalance. For each of the three selected specialties, we use random undersampling to create four class distributions. Random Forest and Logistic Regression learners are built and evaluated based on fraud detection performance. Good fraud detection is demonstrated through the use of random undersampling, across three selected medical specialties. Statistically significant results are seen across the class distributions, with the 80:20 distribution having the best results. Overall, Random Forest (with either 100 or 500 trees), for each class distribution across all specialties, significantly outperforms Logistic Regression, with average AUC scores of 0.881 and 0.88, respectively.