Health care fraud classifiers in practice

Statistical and machine learning methods have become paramount in order to handle large size claims data as part of health care fraud detection frameworks. Among these, predictive methods such as regression and classification algorithms are widely used with labeled data. However, the imbalanced nature of health care claims data and skewness of fraud distributions result with challenges in practical applications. This paper presents the use of various classification algorithms and data pre-processing methods on claim payment populations and overpayment scenarios with different characteristics. It can help the health care practitioners evaluate the advantages and disadvantages of these analytical methods, and choose the right classification method and apply them properly for their specific circumstances. We utilize publicly available U.S. Medicare Part B health care claims payment data from the hospitals with a number of fraud label scenarios to demonstrate potential fraud patterns. We discuss the computational demand and accuracy of the methods.