Health Care Fraud Detection with Community Detection Algorithms

Abstract

Fraud detection is interesting research topic and it not only needs data mining techniques but also needs a lot of inputs from domain experts. In health care claims, relationships between physicians and patients form complex communities structures and these communities could lead to potential fraud discoveries. Traditionally, researchers have focused on clustering physicians and patients and tried to find the suspicious communities. In this paper, we studied and discussed different types of relationships and focus on small but exclusive relationships that are suspicious and may indicate potential health care frauds. We developed two algorithms to detect these small and exclusive communities. These algorithms can be applied to larger dataset and are highly scalable. We tested these algorithms with a set of synthesized datasets. These synthesized datasets were created to resemble the real health care claims datasets and used to test the fraud detection algorithms. The test results show the these algorithms are very efficient and can evaluate the communities structures of 50,000 providers in about 1 minute.