Investigating the effects of class imbalance in learning the claim authorization process in the Brazilian health care market
Jackson Cunha Cassimiro, Andre Macedo Santana, Pedro Santos Neto, Ricardo A. L. Rabêlo · 2017
Fraud and abuse are two factors directly related to high health care costs, since they correspond to expenses that can be eliminated without prejudice to the quality of services provided. In Brazil, the health insurance companies implement a claim authorization process which assists in the detection of fraud and abuse. This process consists of a prior analysis of the services requested by providers, allowing them to detect patterns linked to fraud and abuse. This analysis is commonly performed manually, making the execution expensive and non-scalable. Health insurance companies have invested in the use of data mining and machine learning techniques to detect suspicious fraudulent patterns. However, the use of these techniques in claim authorization process is affected by the class imbalance problem, due to the fact that there are much more authorized service requests than unauthorized ones. This paper presents the investigation results of the effects of class imbalance in health insurance claims authorization domain. By means of an experiment, the performance loss of several classifiers was measured in different class distributions and also the performance recovery provided by treatment methods. The results show that the studied classification algorithms are affected differently by class imbalance. They also show that the recovery performance is lower the higher the class imbalance.