This summary report discusses a project that aimed to provide new policy-relevant applications by using state-of-the-art data science to improve risk assessment of physician engagement in fraud.
The authors of this summary report present the results of a project that sought to answer three research questions: if models that use big data on non-clinical physician behavior, such as patient complaints or life stressors, can successfully predict engagement in fraud in the near-term future, approximately one to five years; which of those non-clinical behavioral factors would represent the greatest risk for fraud engagement; and which machine learning algorithm would be most accurate in predicting a physician’s risk of engaging in fraud. To answer those three questions, the researchers relied on techniques that used behavioral big data and deployed state-of-the-art data analytic tools to detect the risk of Medicare fraud early on, before payment, and potentially early enough to prevent the fraud from occurring. The researchers investigated two kinds of models, predictive and explanatory; their main predictive modeling findings revealed a high degree of performance accuracy up to five years prior to the exclusion year; the explanatory models indicated that factors associated with high risks of fraud included a prior criminal case, tax liens, property purchases, gifts from companies, and one-star online reviews. Both sets of results pointed to the utility of using nonclinical data in fraud prevention and control efforts.