A three-stage framework to detect health insurance fraud.

摘要

Insurance claim denials, fraud, and abuse bring additional expenses to healthcare costs; hence, it is essential to develop a system that can reduce the time and money spent on healthcare insurance claim denials, fraud, and abuse. In this research, a framework for detecting insurance claim denials, fraudulent, and abusive activities in healthcare systems is developed.;The proposed framework comprises three independent stages. The first stage of the methodology analyzes claim denials from provider's perspective by proving that there is a relationship between the overuse of resources and claim denials. Afterwards, it employs distance-based outlier detection and data binning techniques to determine providers who behave differently than their colleagues.;The second stage of the framework considers the relationship among diagnoses, services, and medications on claim forms. The main idea behind this stage is that every group of diagnoses has unique items called services and medications required to treat patients; therefore, combinations of those items that are not necessary for that group of diagnoses indicate suspicious activities. The proposed method applies frequent itemset mining incorporated to multi-objective optimization techniques, in order to extract the relationship among those items. The detection system assumes that claims are suspicious and must raise a red flag if items on claim forms are not on the list of extracted combinations of diagnoses, services, and medications.;The final stage of the methodology analyzes the relationship among patient's demographics, diagnoses, services, and medications prescribed to patients. It is known that some of those items belong to specific groups of the population (e.g. ovarian cancer occurs only in women). The model in this stage utilizes an approach based on risk quantification and decision trees, in order to figure out which items belong to which population groups. Claims are suspicious if the risk value gathered via the risk quantification technique is greater than the threshold obtained from the decision tree.;As a result, the detection system is tested using two data sets, each of which contains claims from providers from different specialties. Accuracy, sensitivity, and specificity rates are employed to measure the performance of the proposed algorithms.