The Algorithm Knows

We fed 10 years of Medicare billing data — every claim, every dollar, every provider — into a machine learning model trained on 8,300+ confirmed fraud convictions. Then we asked it a simple question: Who else bills like a criminal?

The algorithm returned 500 names.

Five hundred active Medicare providers whose billing patterns are statistically indistinguishable from doctors, nurse practitioners, and clinics that were later convicted of healthcare fraud, excluded from federal programs, or sentenced to prison. Collectively, these 500 providers have billed Medicare over $400.0M in taxpayer money — an average of $800.0K each.

This doesn't mean they're guilty. It means the math says they look exactly like people who were.

How We Built It

Our model is a Random Forest classifier — an ensemble of hundreds of decision trees, each one learning slightly different patterns in the data. Think of it as 500 fraud investigators, each with a slightly different perspective, voting on whether a provider's billing looks suspicious. When the majority vote yes, the probability goes up.

The key insight: we don't try to detect fraud directly. We trained the model on the billing patterns of people who were convicted of fraud, then asked it to find active providers who match those patterns. It's the difference between trying to define fraud in the abstract vs. learning what fraud actually looks like from 2,198 real cases.

What the Algorithm Learned

When you train a model on thousands of convicted fraudsters, certain patterns emerge. Here are the features the model found most predictive, ranked by importance:

The #1 predictor surprised us: years active. Fraudsters don't tend to build long careers. They enter the system, bill aggressively for a few years, and either get caught, move to a different NPI, or flee. A provider who's been billing Medicare for 15+ years is statistically less likely to match fraud patterns — not because fraud doesn't exist among veterans, but because the convicted ones rarely lasted that long.

The #2 predictor — services per beneficiary — is the classic fraud signal. Legitimate internists might bill 2–4 services per patient visit. Fraud mills bill 8, 10, 15 services per encounter — stacking labs, tests, and procedures onto every patient who walks through the door.

The Top 10 Flagged Providers

These are the 10 active Medicare providers whose billing patterns most closely match those of convicted fraudsters. Each link goes to their full provider profile with detailed billing data.

Click any provider name to see their full billing history, year-by-year breakdown, and top procedures.

The Specialty Pattern: Why Internal Medicine Dominates

Of the 500 flagged providers, the specialty breakdown is striking:

Internal Medicine + Family Practice = 79.6% of all flagged providers. Why? These are the highest-volume primary care specialties. They see the most patients, order the most tests, and bill the most line items. That volume creates opportunity — more claims mean more chances to pad, upcode, or bill for services never rendered. It's the same reason bank robbers target banks with the most cash.

This doesn't mean internists are inherently more fraudulent. It means the model learned that fraud patterns concentrate where billing volume is highest — which is exactly what DOJ prosecution records show.

The Geography of Suspicion

California and Florida are tied at 56 flagged providers each. Then New York (39), Texas (36), and New Jersey (33). If that list looks familiar, it should — these are the exact same states that dominate DOJ healthcare fraud takedowns year after year.

South Florida alone has been the target of more Medicare fraud strike forces than any other region in the country. Our algorithm didn't know that. It simply learned from the data that providers in these states are more likely to match convicted fraud patterns. The geography of fraud is not random.

The Money

Four hundred million dollars. That's the total Medicare payments flowing to these 500 providers over the past decade. To put that in perspective: the entire HHS OIG budget for investigating Medicare fraud is roughly $400M per year. We're talking about the same amount of money going to just 500 providers whose billing patterns match convicted criminals.

Not all of that money is fraudulent — probably not even most of it. But if even 10% represents waste, fraud, or abuse, that's $40 million in taxpayer money. Enough to fund 800 nurses for a year.

Does It Actually Work?

Yes. Our model achieves an AUC of 0.83 — meaning it correctly distinguishes between fraudsters and legitimate providers 83% of the time. That's not perfect, but it's significantly better than random chance (0.50) and comparable to models used in financial fraud detection.

More importantly: we've already seen it work in the real world. Several providers flagged by earlier versions of our model were subsequently charged by the DOJ — months or years after our analysis identified them. We documented these cases in “Our Data Predicted It”.

The model doesn't know why a provider bills the way they do. It only knows that the pattern matches. But when you train on 2,198 confirmed cases and the math keeps being right, the patterns deserve scrutiny.

Explore the Data Yourself

We believe in transparency. Every number in this article comes from publicly available CMS data. You can verify it, challenge it, or build on it.