Our first Regen AI models learned from about 12,000 CDC health records. The new ones are built on more than 2.6 million.
That's over 200 times more data, and it matters more than it sounds. A risk model learns by studying real people. It sees how age, weight, blood pressure, smoking and family history line up with who ends up having a condition. With 12,000 records, plenty of those combinations showed up only a few times, so the model was working from thin evidence. With 2.6 million, it has seen far more kinds of people, and the patterns hold up much better.
Where the data comes from
Both sources come from the CDC. The first is NHANES, the National Health and Nutrition Examination Survey. We use the 1999 to 2023 data, which covers about 66,000 adults who answered a survey and also had a physical exam and lab tests. It's the gold standard for this kind of work. Nobody has to guess their blood sugar or kidney function, because it was measured.
The second is BRFSS, the Behavioral Risk Factor Surveillance System. We use the 2019 to 2024 data, which covers about 2.6 million adults who answered a CDC health survey by phone. It has no lab work, but it has scale, and scale shows a model how common each condition really is across the country.
Six conditions use both sources combined: heart disease, heart attack, stroke, cancer, arthritis and COPD. The other five use NHANES only: diabetes, high blood pressure, heart failure, kidney disease and fatty liver. Those depend on lab results or questions the phone survey doesn't ask. We tested this split against the alternative and it came out ahead. So the accurate way to say it is that Regen AI is built on 2.6 million+ CDC records, and each condition gets the data that fits it.
What got better
More data gave us more accurate models. We measure accuracy with AUROC, a score from 0 to 1 that shows how well a model separates people who have a condition from people who don't. A score of 0.5 is a coin flip, and closer to 1 is better. We tested on people the models never saw during training, so the numbers reflect how the models do on someone new. Heart disease now scores between 0.90 and 0.92. Heart attack and heart failure land between 0.89 and 0.91, diabetes between 0.85 and 0.86, and stroke between 0.82 and 0.85.
The scores are also more realistic. Early versions ran higher than real life. The average diabetes score was about 30% when the real rate in the population is about 14%. Now the average score for each condition matches how common that condition actually is.
The models also behave in a way that makes sense. Risk can never go down when something that raises risk goes up. Getting older, a higher BMI, smoking, higher blood pressure, a higher A1C or more time sitting will never lower your score. The old models sometimes did, which confused people, and they were right to be confused.
One more thing. Every model can now be rebuilt from scratch from the training code, so any future update can be checked against this one.
If you've used Regen AI before
If your numbers drop on your next assessment, the model change is the most likely reason.
To make that easy to see, your dashboard trend chart now draws a dashed gold line at the point where the model changed. Scores before the line came from an older version and scores after it came from the new one, so you always know what you're comparing.
What Regen AI is for
Regen AI is a free tool. You answer questions about yourself, it estimates your risk for 12 chronic conditions, and it tracks how those estimates change over time on your dashboard. It shows your risk so you know what to bring up with your doctor.
Take the updated assessment at regenai.health/assessment. And if you want updates like this one, create a free account and turn on the weekly newsletter in My Account.
