Medicare banned 99 for fraud.
The model found 918 more.
No fraud labels, no training data. The model grouped 294,740 Medicare doctors by billing behavior and found four clusters with up to 6.6× the banned-doctor rate. The 918 are leads you didn't have.
DisclosureBenchmark study on public CMS.gov data. No customer data was used. Reviewed by Drew Wasem, Founder, Coherany. Methodology available on request.
HOW WE DID IT
We hid the answer key. Then we graded the paper.
Medicare loses $60 to $100 billion a year to fraud. The government publishes a list of banned doctors every month. A name only lands on it after an investigation, and investigations take years. By the time the name shows up, the money is gone. The usual response is to train a model on past fraud cases. That works when your training data is clean, current, and representative. In healthcare fraud, it almost never is.
So we flipped it. We hid the exclusion list entirely. We gave the model nothing but billing behavior: what each doctor prescribes, how much, how often, through how many suppliers, at what markup. It grouped the 294,740 doctors by behavioral similarity with zero knowledge of who had been banned.
Then, and only then, we checked which groups the banned doctors had landed in. Four of them contained excluded doctors at 3.7× to 6.6× the baseline rate, with odds of that being random as low as 1 in 21,000. Every other doctor inside those same groups shares the identical billing signature. They just haven't been investigated yet.
WHAT WE STARTED WITH
Every Medicare doctor who wrote a medical equipment prescription in 2023.
WHAT WE FOUND
Four groups. Four states. Same billing signature as the banned ones.
Baseline across the whole dataset: 1 banned doctor per ~3,000 doctors. A group 'wins' if banned doctors show up in it at 3× that rate with under a 1-in-20 chance of being random. Four groups cleared the bar. All four are family practice doctors in different states with different prescription mixes, but the same billing signature.
California family practice: odds of being random 1 in 21,000
339 California family practice doctors, grouped purely by how they bill. 9 of them are on the federal exclusion list, where the baseline rate would predict 0.12. The probability of that concentration showing up by chance is about 1 in 21,000. Every other doctor in that group (330 of them) shares the same billing signature but hasn't been investigated yet.
Tennessee family practice: 6.6× the baseline rate
122 Tennessee doctors, 4 already banned. That's 3.28% versus the 0.034% baseline, a rate 6.6× higher than random. Different state, different prescription mix from California. Same behavioral archetype. The model had no geography column to anchor on. It found the pattern from billing behavior alone.
The pattern shows up in four states, same specialty
All four winning groups are family practice doctors. Tennessee, California, Pennsylvania, Texas. Different prescription mixes. Same billing signature. That shared signature, not the individual doctors, is the policy object a fraud analyst can review, approve, and run against new claims every month. Approve it once, match it forever.
THE LIST YOU DIDN'T HAVE YESTERDAY
918 doctors bill exactly like the ones Medicare already banned. Nobody has looked at them yet.
The four groups contain 940 doctors. 22 of them are already banned. That means 22% of the known bad actors were sitting inside 3% of the sampled population. The other 918 bill the same way and have never been flagged. That's the queue.
No fraud labels. No training set. No hand-picked features. Under 20 minutes on a laptop, from the raw CMS file to the list.
“Did it find fraud? No. That's the point. It found where to look. The exclusion list was a validation check, not a training target. Everything the model saw was billing behavior.”
HONEST LIMITS
What this does and doesn't prove.
All four winning groups are family practice doctors. We have not yet checked whether family practice physicians are banned from Medicare at a higher rate than the 0.034% population baseline. If they are, some of the measured effect is specialty risk rather than pure billing behavior. The groups are still a valid triage signal: every unknown doctor inside them is a ranked lead for human review. They are not, on their own, evidence that any individual doctor has done anything wrong. The next step in a real deployment is to re-run the enrichment against a specialty-matched baseline and let a fraud analyst adjudicate each lead.
This isn't really about medical equipment.
The same pipeline runs on your claims, enriched with your SIU's prior cases instead of the government's exclusion list. Point it at your book of business and we'll show you where to look. Request the methodology or schedule a walkthrough.