Using PageRank to Detect Anomalies And Fraud in Healthcare
https://hortonworks.com/blog/using-pagerank-detect-anomalies-fraud-healthcare/
This 3-part study, completed by Ofer Mendelevitch and Jiwon Seo Ph.D., takes a closer look into the PageRank algorithm and how it can be used. They set out to see if they could apply the PageRank algorithm to a Medicare dataset for the purpose of detecting fraud and anomalies on a real-world healthcare payment platform.
The writers took a large graph of about 880,000 healthcare providers, where healthcare providers are nodes, and an edge between A and B exists if the two providers are “similar”, which is based on their shared procedure codes (CPT). All specialties, such as dermatology, were then looped so that “source vertices” included all medical providers with that specialty. Then the PageRank algorithm was applied to the source vertices. The resulting scores were used as specialty-specific PageRank scores for a given node to detect anomalies. For example, if a PageRank score for Dermatology is computed, and a doctor in that group specializes in surgery but has a relatively high PageRank score, then he is identified as an anomaly. The study resulted in having about 446,000 nodes and about 673 million edges.
It is almost bizarre to see a topic learned in class being applied to solve real-world issues. One would never think about taking a technological approach, instead of working through say, law enforcement or the government, to detect healthcare fraud. As impressive as it is for the authors to apply this algorithm to a very difficult, real-world problem, there are some things I believe they did not take into account. If a provider has multiple specialties, which is possible in the real world, then the algorithm would flag him/her as an anomaly, even though they are not fraudulent. However, despite the slight inaccuracies, this study provides an important model for others follow, and maybe one day in the future PageRank will be used to solve lots of other problems in the world.