Medical Provider Specialty Predictions for the Detection of Anomalous Medicare Insurance Claims

Abstract

Fraud, waste, and abuse in medical insurance contributes to significant increases in costs for providers and patients. One way to reduce costs is through the detection of abnormal medical practices that could indicate possible fraud. In this paper, we expand upon our previous research into medical specialty anomaly detection by validating the efficacy of our model using real-world fraud cases, and then testing three strategies to improve model performance. The three strategies are feature selection (to include adjusting for class imbalance), medical specialty grouping, and the removal of specific, overlapping specialties. We use the publicly available Medicare claims data, released by the Center for Medicare and Medicaid Services, for building and testing our models. In addition to using the 2013 data, we use the 2014 data for model validation and comparisons. We employ the List of Excluded Individuals and Entities (LEIE) database, released by the Office of Inspector General, as well as two other documented fraud cases, for model testing. Multinomial Naïve Bayes is used to build all models. In this work, we confirm our prior model was able to correctly classify 67% of the real-world fraudulent physicians contained in the LEIE database as fraudulent. Furthermore, the three proposed strategies show good results in improving model performance.