An Application of Exploratory Data Analysis in Healthcare Fraud Detection.

• In Event: Breakfast with Research Roundtable Forum Session
  In Paper Session: Table 1

Abstract

Health care has become a major expenditure in the US since 1980. Therefore, effective fraud detection is important for reducing the cost of U.S. health care system.numerous researchers have attempted to develop more sophisticated antifraud approaches incorporating data mining, machine learning or other methods (Chan and Lan, 2001; He et al., 1997, 2009; Liou et al., 2008; Major and Riedinger, 2002; Musal, 2010; Otega et al., 2006, Sokol et al., 2001; Viveros et al., 1996; Williams and Huang, 1997; Yamanishi et al, 2004). Compared with existing fraud detection system, these newly proposed approaches focus on more complicated tasks such as automatic learning of fraud patterns from data, prioritizing suspicious cases by assigning “fraud likelihood” to each case, and identifying new types of fraud which were not previously documented. Due to their exploratory nature, most of these methods can be considered as exploratory data analysis (EDA) techniques, focusing on identifying trends, patterns and outliers from the data (Tukey, 1977).
However, these studies are essentially technical in character. They mainly focus on how to develop various methods to identify different types of fraudulent behaviors in healthcare practice from computer science or information technology perspective. Few of them discuss healthcare fraud detection from accounting/auditing point of view. Since one of the responsibilities of both internal and external auditors is fraud detection, how auditors can integrate these healthcare fraud detection techniques in audit process and how they can benefit from these methods is worth studying. Therefore, this study intends to fill this gap and demonstrate how auditors can take advantage of EDA techniques to assess healthcare fraud risk by following the conceptual EDA application process proposed by Liu (2014), which include eight normative steps: display the distribution of fields related to audit object, identify salient features, generate possible explanations, test the explanations, identify suspicious case, explore the causes of abnormal cases, confirm the relationship, add new audit objectives and report the results. Specifically, a real Medicare inpatient claim dataset purchased from Center for Medicare and Medicaid Services (CMS) is used in this case study. Moreover, EDA techniques such as descriptive statistics, data visualization, cluster analysis and association analysis are applied. To reveal the benefits auditors can gain from EDA approach, EDA results are compared with the results generated from conventional audit procedures.

Authors

• Qi Liu, University of Rhode Island

• Miklos A Vasarhelyi, Rutgers University - Newark