SBIR Phase I: Predicting Healthcare Fraud, Waste and Abuse by Automatically Discovering Social Networks in Health Insurance Claims Data through Machine Learning

点击翻译按钮获取中文摘要

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will pave the way for new types of social network analysis to detect anomalies, which could lead to more accurate and faster identification of Fraud, Waste, and Abuse (FWA), key opinion leaders (i.e., influentials), and market segments. Medicare and other healthcare providers lose hundreds of millions of dollars to FWA. This research proposes using a novel way to discover and combine relationships between entities (e.g., doctors) with information about the entities (e.g., prescription history) using machine learning. The goal is to reduce a claims investigator's workload while maintaining high accuracy in detecting FWA. In short, the results of this research will not only improve FWA detection efficiency, but enable detecting new types of FWA. Societal impact includes reduced costs to the taxpayer for government supported programs such as Medicare through better FWA detection. More broadly, the system could be used to find terrorist and crime networks, detect possible opioid or substance abuse epidemic cohorts, under-medication, over-medication, and even incorrect medications.The proposed project will apply a novel machine learning method to solve the Fraud, Waste, and Abuse (FWA) problem in health insurance. The technical problem is how to combine relations between entities such as doctors with information about doctors (e.g., a doctor's prescription history). This project advances the state of the art by developing a new way to automatically discover those relations and then combining those relations with the information about doctors through machine learning, thus vastly improving prediction accuracy. The method uses relation information to fill in the gaps of entity information alone and vice versa. It is believed that this method will hugely improve the ability to detect FWA. The goal is to achieve a 50% true positive rate in a database of fraud-convicted doctors published monthly by the government. The scope of the project involves analyzing several different types of health insurance claims formats (e.g., Medicare) and producing a fraud score, which then others can use. The anticipated results include a fraud score for most doctors in the U.S. (at least those who deal with Medicare), APIs to these scores, and an interactive visual system that claims investigators can use to reduce their workload while accurately identifying FWA.