Bayesian Methods for Assessment of Health Care Fraud

Health care fraud costs a significant amount to the US healthcare system. According to FBI the fraud cost to US is about $80 billion a year. Detection and prevention of fraud are of interest to many parties including government, medical community and insurance industry. Due to complexities involved in defining fraudulent behavior and detecting fraud, the need for more sophisticated data mining tools and statistical methods have been recognized in the health care industry. As a result, methodologies such as neural networks, decision trees, association rules, link analysis and genetic algorithms have been used in for fraud detection. However, use of Bayesian methods has been quite limited with the exception of some recent work using Bayes networks.
In this talk we consider Bayesian methods for fraud detection. In so doing, we focus on co-clustering models for analysis of dyadic data connecting two entities such as providers and beneficiaries. These models can be used to identify billing patterns of potentially fraudulent providers who submit a large number of high-dollar claims to insurance programs. An important issue pertinent to dyadic data is to identify the cluster membership of each entity. We discuss how Bayesian inference can be developed for assessment of cluster memberships and illustrate what additional insights can be obtained from Bayesian analysis of the dyadic data. We consider different forms of dyadic data and estimation of cluster sizes and discuss potential extensions such as dynamic cluster membership.