Identification of Misclassified Medicaid Audits

Abstract

Medicaid auditors typically use statistically valid samples from the total claims filed by service providers to detect overpayments or underpayments. These samples often contain many "zeros" (compliant claims) and a few "ones" (flagged issues), but errors in the auditing process can occur due to human mistakes, biases, or even AI and machine learning errors. Such mistakes may lead to false-positive results where a claim is wrongly flagged as an overpayment. Is it possible to identify and isolate false positives in defending impugned claims filed by Medicaid providers without the need for expensive forensic audits of the entire sample? To address this, we test the effectiveness of noise-filtering algorithms to isolate false-positive claims from legitimate ones. By artificially injecting noise (representing false positives) into synthetically generated data we create a realistic litigation environment resembling what happens during subpoenas or document requests. We then apply three noise-filtering algorithms and find that these filters reduce the audit data to a smaller, more focused sample, making it easier to identify false positives during any subsequent manual review. Our study does not propose a novel noise-filtering method; rather, we demonstrate how existing techniques can help forensic analysts concentrate false-positive claims in a reduced sample. While the specific focus is on Medicaid audits, the findings are applicable to any situation where audits are conducted on claims filed by service providers and paid by third parties.