An interactive machine-learning-based electronic fraud and abuse detection system in healthcare insurance

• The nature of problem and claim management environment are well-defined.
• Very comprehensive literature revive is presented.
• An IML based novel DSS detecting fraud and abuse in healthcare insurance is developed.
• Unlike earlier studies, the system embraces all relevant actors and commodities.
• Real data are used for the evaluation and perform better results wrt previous studies.

Detecting fraudulent and abusive cases in healthcare is one of the most challenging problems for data mining studies. However, most of the existing studies have a shortage of real data for analysis and focus on a very limited version of the problem by covering only a specific actor, healthcare service, or disease. The purpose of this study is to implement and evaluate a novel framework to detect fraudulent and abusive cases independently from the actors and commodities involved in the claims and an extensible structure to introduce new fraud and abuse types. Interactive machine learning that allows incorporating expert knowledge in an unsupervised setting is utilized to detect fraud and abusive cases in healthcare. In order to increase the accuracy of the framework, several well-known methods are utilized, such as the pairwise comparison method of analytic hierarchical processing (AHP) for weighting the actors and attributes, expectation maximization (EM) for clustering similar actors, two-stage data warehousing for proactive risk calculations, visualization tools for effective analyzing, and z-score and standardization in order to calculate the risks. The experts are involved in all phases of the study and produce six different abnormal behavior types using storyboards. The proposed framework is evaluated with real-life data for six different abnormal behavior types for prescriptions by covering all relevant actors and commodities. The Area Under the Curve (AUC) values are presented for each experiment. Moreover, a cost-saving model is also presented. The developed framework, i.e., the eFAD suite, is actor- and commodity-independent, configurable (i.e., easily adaptable in the dynamic environment of fraud and abusive behaviors), and effectively handles the fragmented nature of abnormal behaviors. The proposed framework combines both proactive and retrospective analysis with an enhanced visualization tool that significantly reduces the time requirements for the fact-finding process after the eFAD detects risky claims. This system is utilized by a company to produce monthly reports that include abnormal behaviors to be evaluated by the insurance company.

Figure optionsDownload as PowerPoint slide