Prospective vs. Retrospective Risk Adjustment, and RADV Audit Exposure

The choice is defensibility, not timing

On paper, prospective and retrospective capture differ only in when the coding happens. Prospective capture starts before the visit, when the plan mines prior claims and clinical data for members whose chronic conditions look under-documented, flags those gaps for the physician, and confirms that whatever she treats is documented and coded that day. Retrospective capture starts after the claim, when a reviewer reads the closed chart weeks or months later, hunts for diagnoses the claim missed, and files them through chart review. In a disciplined plan, the two operate as halves of one quality process.

The difference matters most when an auditor arrives. A prospective program hands her a clinical note that the physician wrote while delivering care, with the reason for the diagnosis already in it; a retrospective program hands her a billing code that a reviewer proposed after the plan was paid, reading specifically for what the claim missed. The reviewer, tellingly, has no reason to look for codes to remove: the plan pays him to find the ones that add revenue, so his search runs in a single direction.

OIG has measured how far that one-directional search goes. In a 2017 review, it found that MA companies collected an estimated $9.2 billion from diagnoses that surfaced only in chart reviews and health risk assessments, with no other record that the care ever happened; the money concentrated sharply, as just 20 of 162 companies claimed 54% of it while covering only 31% of members. None of this indicts chart review itself, but a program that only ever adds codes, and never deletes one, has stopped reviewing and started inflating the score.

AI changes the speed, not the standard

AI has moved into both ends of this workflow, and plans routinely misjudge what it changes. Before the visit, a model reads a patient’s record and flags conditions for the physician to check; after the visit, another model reads the finished chart and proposes codes for a coder to confirm. Both make the work faster, and neither can tell whether the record actually supports the code it surfaced. Everything turns on that distinction: a model that hands the physician a condition to evaluate helps her document care as she delivers it, while a model that drops a code onto a claim with no physician behind it only inflates the score faster. The auditor still asks the same question, and AI decides only how much of it you will have to answer for.

RADV puts a price on the gap

Historically, RADV carried little financial force, because CMS recovered overpayments one patient at a time and rarely collected sums that changed a plan’s behavior. The 2023 Final Rule, CMS-4185-F2, shifted that scale by letting the agency audit a sample of a contract’s patients and apply the error rate it finds to the entire contract, which turns a modest sampling result into an enormous number: on a $1 billion contract, a 5% error rate extrapolates to a $50 million bill, and CMS expected to collect roughly $479 million a year once the audits reached payment year 2018.

Compounding this, the same rule removed the FFS Adjuster, and that removal acts as a multiplier on the extrapolation penalty. CMS had built the adjuster in 2012 to fix a fairness problem: the risk model is calibrated on traditional Medicare’s own claims, which carry their own unsupported diagnoses, so auditing a plan against a spotless standard would hold it to a bar the government does not meet itself. The adjuster discounted RADV recoveries to account for that gap, so stripping it out means CMS no longer offsets a plan’s errors against the errors baked into fee-for-service data. Extrapolation multiplies the sample across the whole contract; losing the adjuster multiplies the result again.

What makes those bills collectible, rather than theoretical, is that the errors keep turning up. Across OIG’s high-risk reviews, about 70% of the sampled codes had no support in the record, and some categories failed more than 90% of the time. In one audit, OIG pulled the charts for 97 patients whose plans had billed Medicare for an acute stroke and found that not one documented a stroke, a result it projected to $462 million in overpayments to MA plans for 2021 on stroke coding alone. CMS did not invent this exposure; the plans’ own documentation did.

The 2025 ruling is a narrower reprieve than it looks

That calculus appeared to collapse in September 2025, when a federal judge in Humana v. Becerra vacated CMS-4185-F2; CMS has appealed. Many plans read the decision as a reprieve, but it is far narrower than that: the judge never ruled that extrapolation is illegal or that the adjuster must return. He found instead that CMS had changed its legal reasoning between the 2018 proposal and the 2023 rule without opening the new version to comment, so the rule failed on process rather than on the merits, a defect CMS can cure by proposing it again and running it cleanly, and one the appeal could undo outright. Ultimately, the ruling lifts the machinery that made errors expensive, the extrapolation and the missing adjuster alike, but it leaves the underlying error rate exactly where it was, and it does nothing to stop OIG, patient-level recoveries, or the shift to the V28 risk model that is already reshaping how much risk score a plan can defend. A plan counting on the appeal has chosen to fix a documentation problem in court.

What holds, whatever the court does

If the exposure really depends on which end of the payment a plan codes at, then the answer is not to wait out the appeal but to move the work earlier. The plans that will hold up capture the diagnosis at the visit, where the physician’s note already carries its own support, and they treat retrospective review as genuine quality control, judged as much by the codes it removes as the ones it adds; a review that has never deleted an unsupported diagnosis is quietly building liability with every cycle. For every risk-adjusting code, they keep the exact line in the chart that backs it, ready to show an auditor the day he calls, and they hold their software to the same standard they hold their coders: it may suggest and document, but it may never bill on its own. None of this turns on the appeal, because the law may swing on extrapolation for years while the auditor’s question stays fixed: does the chart support the code?

Sources

• CMS, “Medicare Advantage Risk Adjustment Data Validation Final Rule (CMS-4185-F2) Fact Sheet.” cms.gov
• Federal Register, “Medicare and Medicaid Programs: Policy and Technical Changes ... (RADV),” Feb 1, 2023 (extrapolation from PY2018; ~$479M/year estimate). federalregister.gov
• OIG, “CMS Potentially Overpaid Medicare Advantage Organizations $462 Million Based on Certain Unsupported Acute Stroke Diagnosis Codes.” oig.hhs.gov
• OIG, “Toolkit to Help Decrease Improper Payments in Medicare Advantage Through the Identification of High-Risk Diagnosis Codes” (~70% unsupported; some over 90%). oig.hhs.gov
• OIG, “Billions in Estimated Medicare Advantage Payments From Diagnoses Reported Only on Health Risk Assessments and Chart Reviews” (OEI-03-17-00474; the $9.2B finding). oig.hhs.gov
• Milliman, “Federal court vacates 2023 rule on CMS RADV audits,” with Groom Law Group and Crowell & Moring analyses of Humana Inc. v. Becerra (Sept 25, 2025 vacatur; CMS appeal). milliman.com

Figures are drawn from CMS, the Federal Register, and OIG primary sources, with the September 2025 vacatur and CMS appeal confirmed across independent legal analyses. Workflow descriptions reflect standard industry practice. The RADV Final Rule is in active litigation; confirm the current posture before relying on its extrapolation provisions.