James
In December 2023 a US federal watchdog did something regulators almost never do. It published exactly how it catches health insurers billing for diagnoses that patient records do not back up. Not a policy statement. The actual database queries.
Two and a half years later it ran that method across the country. Ninety-seven patient records, all billed for an acute stroke. Everyone failed. Estimated cost to taxpayers for a single year: about $462 million.
The answer key had been public the whole time.
How the money works
Medicare Advantage is the privatised branch of US Medicare. The government pays insurers a fixed monthly sum per patient, more for sicker patients. That design is deliberate and sensible. Without it, insurers would dodge anyone expensive.
Sickness gets established through diagnosis codes pulled from medical notes. Sicker diagnoses, bigger payments. So what a doctor writes, and how someone later codes it, moves real money.
It is not the scandal it looks like
“97 out of 97 failed” sounds like organised fraud. The detail says otherwise, and the detail is more interesting.
In 68 of those cases the patient had genuinely had a stroke. Documented, clear, right there in the record. The problem was when. The stroke had happened earlier, and the insurer billed a code describing it as a current emergency.
That distinction is not wordplay. A stroke happening now and a stroke from three years ago are different codes worth different amounts. Billing the first when the record supports the second is an overpayment, whatever anyone intended.
The rest were more ordinary: 22 records with no stroke documentation at all, four that nobody could find, one signed by a pharmacist instead of a physician, one that was simply illegible.
The trick behind the detection
The published method does not try to second-guess doctors. It hunts for contradictions.
A real acute stroke leaves a trail: an ambulance, an emergency department, an admission, a brain scan. So the query asks one question. Is there an acute stroke on a doctor’s office record, with no matching hospital record anywhere that year?
When those two facts disagree, a human opens the chart. That is the entire idea.
The watchdog had already published what it kept finding: “approximately 70 percent of those diagnosis codes were not supported in the associated medical records,” some categories failing “over 90 percent of the time.”
Why hand over the method?
Regulators normally guard this stuff. The reasoning here was practical. Organisations had been asking how the selections were made, and the watchdog decided more of them fixing their own errors beat catching them later.
So: an open-book exam, questions released in advance.
The 2026 results suggest most did not open the book. That is less about dishonesty than about what got built. Software for reviewing medical records was designed to find diagnoses that had been missed, because finding more raised revenue. Software designed to find things that should not be there is a different product entirely, and far fewer organisations ever bought one.
The bit that is not about healthcare
Any system that flags things automatically gets built around whatever its owner is measured on. Loan applications, insurance claims, job applications, all of it. If the measure is finding more of something, that is what it finds.
Asking it to also catch what should not be there is a genuinely different job, needing a different design, and it rarely gets built until somebody outside forces the question.
Here the force was a regulator publishing its own method and then demonstrating what it catches. Organisations now investing in a risk adjustment solution that can remove unsupported diagnoses as well as find missing ones are responding to a fairly blunt lesson: the standard was public, the failure rate was documented, and one year of one condition came to $462 million.
