The Anatomy of an Unsupported Diagnosis: What the OIG Audits Reveal
In 2015, a physician coded an acute stroke on a single office visit.
The chart told a different story. The stroke had happened in 2001 — fourteen years earlier. What the record actually documented was a history of stroke, which carries no risk-adjustment weight at all. The acute code, which does, had no admission behind it, no imaging, no acute event. Just a past tense dressed up as a present one.
That single enrollee-year wasn't an outlier. In the same audit, the plan had submitted acute stroke codes for 54 members. Fifty-two of them didn't hold up.
We recently did something most risk adjustment teams never have time to do: we read OIG's published library of Medicare Advantage compliance audits end to end, alongside OIG's own high-risk diagnosis toolkit. We expected a long list of unrelated coding mistakes.
Instead, we found a pattern — and the pattern is the story.
Across plans, states, and payment years, the same shapes recur. The same handful of conditions, flagged the same way, failing for the same reasons. In some high-risk groups, the failure was nearly total — 30 of 30 sampled records unsupported, not one surviving review. These weren't obscure edge cases. They were predictable, visible in principle in data the plans already had.
That predictability is the whole point. It means an unsupported diagnosis has an anatomy. Learn to recognize it, and you can find these issues in your own data long before an auditor does.
The Anatomy of an Unsupported Diagnosis
Loud vs. Silent
A real condition leaves a trail. A miscode leaves one lonely claim.
A real condition
A genuine acute stroke throws off evidence everywhere:
- Emergency admission
- Brain imaging
- A cascade of follow-up care
- Corroborating claims & pharmacy
- A trail across every data source
The record agrees with itself.
Corroboration is everywhere you look.
A miscoded one
The same code appears once — and then nothing follows:
- A single physician claim
- No admission
- No treatment
- No medication
- No supporting footprint anywhere
The absence is the signal.
The dog that didn't bark.
The principle: presence vs. corroboration
Here is the single idea that unifies every audit we read.
Auditors do not primarily ask whether a diagnosis exists. They ask whether the rest of the clinical record agrees with it.
A genuine medical condition is loud. A real acute stroke generates an emergency admission, brain imaging, a cascade of care. A real cancer generates treatment. A real embolism generates a prescription. These events leave a corroborating trail across claims, encounters, and pharmacy data.
A miscoded diagnosis is silent. It appears once, on a single claim, and then… nothing. No admission where one should be. No treatment. No medication.
The auditor's whole skill is listening for that silence.
It's the clinical equivalent of the dog that didn't bark. The absence is the signal. And it's a signal any plan can learn to hear.
The four questions
Here's the exercise. You have one member, one claim, and no other context. What do you check?
Strip the high-risk condition groups down to their logic, and it comes down to four questions. This is the anatomy itself.
Where's the hospital stay?
Acute stroke, heart attack, and sepsis force a hospitalization. When one is coded on a single office visit with no hospital claim, the record usually describes a history of the event — which carries no risk-adjustment weight — not an active one.
Often dispositiveWhere's the treatment?
Active cancer is treated — surgery, radiation, chemotherapy — and treatment leaves a trail. No treatment in the surrounding months often points to “history of cancer” rather than active disease. But a quiet record here doesn’t always mean an error.
Needs a clinicianWhere's the medication?
Some conditions are defined by how they’re managed. A real embolism is anticoagulated; major depression is typically treated with antidepressants. When the corroborating prescription never appears, the diagnosis and the pharmacy record disagree — and that’s where the finding lives.
Signal strength variesDoes this code belong here?
The quietest of all. A single transposed digit can move a member into an unrelated — and often lucrative — condition they never had. No clinician would catch it, because the mistake is in the data entry, not the medicine. Only a data pattern reveals it.
Often dispositiveHere's where most of this advice goes wrong
Now the part that separates real expertise from a blunt instrument — and the part of our review we'd most want a risk adjustment leader to take away.
The Anatomy of an Unsupported Diagnosis
Not Every Silence Is Guilt
The same missing evidence means different things. Weight the signal before you act.
Absence is near-proof. Act on it.
A real one leaves an unmistakable trail — when it's missing, the diagnosis is very likely a miscode.
Acute stroke
forces admission
Acute MI
forces admission
Sepsis
forces admission
Transposition error
data pattern, not medicine
Only suspicion. Clinician first.
A legitimate, active condition can be quiet — surveillance, hormone therapy, or care out of network.
Prostate cancer
watchful waiting is standard
Breast cancer
hormone therapy alone
Out-of-network care
no Medicare claim generated
Not all of these signals carry the same weight.
Some silences are close to decisive. A genuine acute stroke nearly always produces a hospitalization; there is no common, legitimate scenario where a real one leaves no hospital trail. When that corroboration is missing, the diagnosis is very likely a miscode. The same tends to hold for acute MI, sepsis, and the transposition errors. These are dispositive — the absence is strong enough to act on.
Other silences are only suspicion.
Consider prostate cancer. Active surveillance and watchful waiting are standard, appropriate care for many patients — which means a legitimately coded, genuinely active cancer can have no treatment in the record and look exactly like a miscode. This isn't hypothetical: in the audits, plans pushed back hard on precisely these cases — a member who elected watchful waiting with PSA monitoring, another treated through the VA where no Medicare claim was ever generated. The diagnosis was real. The silence was innocent. The same can be true for a breast cancer patient on hormone therapy alone. These signals are heuristic — a valuable lead, but not proof.
The distinction matters enormously in practice. Treat a heuristic signal as if it were dispositive, and you strip valid diagnoses off patients who genuinely have the condition — creating errors in the opposite direction. Accuracy runs both ways. The goal is never to delete the most codes; it's to arrive at the correct record. That's why the strongest signals can be acted on quickly, while the softer ones belong in front of a clinician before any conclusion is drawn.
Why this outlasts the model
CMS is transitioning its risk model — V22 to V24 to V28 — and each version renumbers, regroups, and reweights the HCCs. It's tempting to think that makes this kind of analysis a moving target.
It doesn't. The clinical corroboration never changes. A real stroke still forces an admission under every model. A real cancer still gets treated. What shifts is only which HCC a diagnosis maps to and what it's worth — the payment layer, not the logic. The questions endure; only the answer key gets renumbered. Build your validation on the clinical logic, and it survives every model change CMS makes.
Why "eventually" is expensive
If these patterns are so knowable, why do they keep surfacing in audits rather than getting caught first?
Rarely for lack of data. Almost always for lack of assembly.
Every one of these four questions requires information from more than one place at once. The diagnosis lives in one system, the missing hospital stay in another, the absent prescription in a third. The finding hides in the space between your data sources — precisely where fragmented systems can't look.
You cannot hear the silence if the record is scattered across three rooms.
And the cost of not hearing it compounds. RADV findings aren't simply counted and corrected; they're extrapolated — a sample projected across the whole population. Go back to that plan with 52 unsupported stroke codes: once the findings were projected across the sampling frame, the estimated overpayment landed at roughly $3.58 million. A handful of quiet coding slips became a multimillion-dollar, contract-level financial event.
Finding it first
The most sobering takeaway from reading these audits is also the most empowering: the findings were audible the entire time. Nothing in them required information the plan didn't already possess. It only required assembling that information and knowing which questions to ask.
That's the shift that separates plans who scramble when an audit notice arrives from plans who are simply, quietly ready — bringing claims, encounters, pharmacy, and chart-review data into a single view, and asking these four questions continuously, so an unsupported diagnosis surfaces before submission rather than after a finding.
The auditors have shown their method. It's remarkably consistent, and it's entirely learnable.
The only real question left is who gets to your findings first — you, or them.