Medicare Fraud Detection Using Random Forest with Class Imbalanced Big Data

Abstract

The healthcare industry generates a substantial amount of data. This big data includes information such as patient records and provider payments. The use of big data is often considered the best way to produce effective models in areas such as fraud detection. In this study, we demonstrate that the use of more highly imbalanced big data does not produce acceptable fraud detection results. We use random undersampling to generate seven different class distributions and compare performance results. We use the 2012- 2015 Calendar Year Medicare Provider Utilization and Payment Data mapping actual fraud labels from the List of Excluded Individuals/Entities (LEIE). Our results, based on building Random Forest models using 5-fold cross-validation, demonstrate that 90:10 is the best class distribution with an 0.87302 AUC, whereas the balanced and two highly imbalanced distributions produced the worst fraud detection performance. Furthermore, we show that the commonly used ratio of 50:50 (balanced) was not significantly better than using a 99:1 (imbalanced) class distribution. Our study clearly demonstrates the need to apply at least some sampling to big data with class imbalance and suggests the 50:50 class distribution does not produce the best Medicare fraud detection results.