Author: Cheryl Ericson, RN, MS, CCDS, CDIP | July 7, 2026
Claims data that deviates from prior patterns or industry trends has been used to identify at-risk claims even before artificial intelligence (AI) entered the revenue cycle.
The Centers for Medicare & Medicaid Services (CMS) created the Program for Evaluating Payment Patterns Electronic Report (PEPPER) in 2002 to summarize provider-specific traditional Medicare data for target areas associated with improper payments due to billing, Diagnosis-Related Group (DRG) coding, and/or admission necessity issues. PEPPER is one of the unique resources that allows hospitals to compare their performances to those of other hospitals within their state, Medicare Administrative Contractor (MAC) jurisdiction, and the nation.
Whereas PEPPER is a resource for hospitals (it requires a secure portal for access), payers can easily access similar data using historical Medicare claims data (MedPar) or their own proprietary data set. A concept that was once used as a compliance tool has been weaponized by payers.
PEPPER uses the top and bottom 20th percentiles to identify outliers. Hospitals that rank within the top 80th percentile are classified as high outliers, compared to their cohorts within a target area. Those in the bottom 20th percentile are low outliers. The U.S. Department of Health and Human Services (HHS) Office of Inspector General (OIG) encourages hospitals to implement a compliance program that includes conducting regular audits to validate that Medicare is being billed correctly. Reviewing a sample of cases within any target area where the facility is a high outlier supports compliant Medicare billing, because these cases are at risk of being audited or denied. Conversely, being a low outlier may indicate revenue opportunities associated with underbilling.
As payers increasingly turn to AI for claims processing, they are gaining insight into billing trends and can more easily and readily identify outliers. An example of the power of modern data analytics emerged in 2018. A data analytic firm, Integra, brought a whistleblower case against two hospitals alleging upcoding. The firm was concerned that these facilities were over-reporting severe malnutrition, encephalopathy, and respiratory failure. Integra made their assertion because these clients deviated from CMS claims regarding the reporting of these major complications and comorbidities (MCCs) over a six-year period. The complaint was based primarily on a statistical analysis of Medicare claims data that demonstrated one of the hospitals submitted proportionally more claims with higher-paying diagnosis codes than comparable institutions. Claims with an MCC rate “more than twice the national rate” or “three percentage points higher than in the other hospitals” were flagged as fraudulent.
Ultimately, the lawsuits were dismissed. The judge ruled that educating physicians documenting using terms supporting the reporting of diagnoses classified as CCs or MCCs was “not in and of itself one to submit false claims.” The education was also determined to be consistent with efforts to improve hospital revenue through “accurate coding of patient diagnoses in a way that will be appropriately recognized and reimbursed by CMS.” Furthermore, the judge found that the medical records did not contain information that was “not justified” by physician judgement and medical opinion. The judge felt that the allegations of fraud overlooked an “alternative hypothesis:” that the hospital was “simply better than their peers in their efforts to ensure their medical documentation and coding maximized the opportunities for legitimate reimbursement from CMS.” It is reassuring that a judge did not jump to the conclusion of fraud, yet many payers are exploiting similar trends as a reason to remove impactful diagnoses from a claim, resulting in a DRG downgrade. This statistical outlier logic that once took years for a whistleblower to compute now happens instantly, at scale, via AI.
Diagnoses that have historically been targets for medical necessity denials or MS-DRG downgrades are more easily detected using AI tools that use past trends (e.g., diagnoses associated with denials) to predict future at-risk claims. The Office of the National Coordinator for Health Information Technology noted that 71 percent of hospitals reported using predictive AI integrated within their electronic medical record (EMR) in 2024. They also found billing to be one of the fastest-growing reasons hospitals are investing in predictive AI. The question is, when these types of claims are identified by the hospital, are they auditing these cases to validate accurate coding and billing?
These tools are becoming more common for both payers and hospitals. FinThrive designed one to help prevent denials by learning “from real-world adjudication outcomes from billions of institutional and professional claims.” This is a different approach than relying on payer policies. As discussed by Frank Cohen in his ICD-10monitor series about AI, payers are increasingly looking for pattern deviations among facilities. Cohen points out, “the audit that used to start with a single flagged chart now starts with a statistical model that has noticed something about your organization’s data. By the time a human auditor opens a chart, the suspicion has already been generated by an algorithm.”
CMS has acknowledged that there is nothing “inappropriate, unethical or otherwise wrong with hospitals taking full advantage of coding opportunities to maximize Medicare payment that is supported by documentation in the medical record.”[1] AI is a double-edged sword. It can identify opportunities for hospitals that can also be used as evidence of potential wrongdoing by payers. I recently highlighted a Blue Cross Blue Shield (BCBS) article in which the payer assumed that hospitals were upcoding acute blood loss anemia during admissions for delivery. Their analysis found that the increased volume of claims coded with acute posthemorrhagic anemia at the analyzed hospitals added $22 million to maternity admission costs in one year. However, their conclusion was biased by their limited definition of “appropriate treatment.” If the claim did not also have a procedure code for a blood transfusion, the diagnosis was considered invalid – yet blood transfusions are not considered a first-line treatment for this condition.
What should clinical documentation integrity (CDI) departments be doing differently? Do not blindly accept denials. An individual claim may be accurately billed, but associated with an at-risk trend. Business as usual is likely to result in more denials, since AI is trained on historical data. The solution:
- Review a sample of historically at-risk claims with your clinical revenue cycle team (CDI physician advisor, denial management team, coding leadership, utilization review leadership, CDI leadership, quality department leadership, and physician leadership);
- Determine how to best leverage Medicare resources like PEPPER and/or Comprehensive Error Rate Testing (CERT) results;
- Identify organizational denial trends, comparing overturned denials with those that were upheld. Address deficiencies associated with claims for which the denial was upheld;and
- Educate staff and adjust workflows as needed to address deficiencies during the concurrent review process to promote efficiency and reduce future rework.
Remember, AI is a tool to identify opportunities (or deficiencies), but keep the human in the loop to validate these recommendations and ensure that the context is accurate.
[1] Medicare Program; Changes to the Hospital Inpatient Prospective Payment Systems and Fiscal Year 2008 Rates, 72 Fed. Reg. 47,130, 47,180 (Aug. 22, 2007).
This article was originally published on RACmonitor.
