Diabetes develops silently over years, giving neither patients nor doctors clear warning signs
In the long struggle against chronic disease, medicine has often arrived too late — treating what might have been prevented. A team of researchers, drawing on the health records of more than three million people, has now built a machine-learning model that may shift that timeline dramatically, identifying individuals at risk for type 2 diabetes up to a decade before the disease takes hold. Presented at the American Diabetes Association's annual conference in New Orleans, the work raises a quiet but profound question: what becomes possible when we can see illness coming from far enough away to change its course?
- More than 60 percent of American adults carry diabetes risk factors, yet prevention programs reach only a fraction — a gap that costs lives and billions of dollars each year.
- The disease advances silently for years, and current screening tools lack the precision to distinguish who truly needs intervention from the vast pool of those merely at risk.
- Trained on 3.36 million Kaiser Permanente patients and combining clinical data with neighborhood-level factors like food access and walkability, the model predicts diabetes risk at one, three, and ten years with striking accuracy — 74% sensitivity and 82% specificity.
- The model outperforms traditional screening by targeting those most likely to benefit, including people who don't fit conventional risk profiles but face genuine danger.
- Researchers are now planning clinical trials to answer the harder question: whether knowing one's risk actually changes behavior — and whether changed behavior actually prevents disease.
Researchers have developed a machine-learning model capable of predicting who will develop type 2 diabetes up to ten years before symptoms appear, using electronic health records from more than three million patients. The findings were presented at the American Diabetes Association's annual scientific conference in New Orleans.
The scale of the challenge the model addresses is immense. Over 60 percent of American adults carry diabetes risk factors, yet prevention programs reach only a small fraction of them. The disease typically develops in silence, offering neither patients nor clinicians clear warning until meaningful harm has already occurred.
The study followed 3.36 million adults treated at Kaiser Permanente Northern California between 2012 and 2024. Using a technique called hazard-based super learning — which blends multiple survival-analysis models — researchers estimated each person's diabetes risk over one, three, and ten years. Inputs included routine clinical data such as age, weight, and blood glucose, as well as publicly available information about neighborhood food access and walkability.
The results were notably precise. Tested on unseen data, the model achieved an AUC score of 0.883, correctly identifying 74 percent of those who would develop diabetes while avoiding false positives in 82 percent of those who would not. Its one-year predictions were nearly exact: a predicted rate of 1.03 percent against an actual rate of 1.01 percent.
Lead author Luis Rodriguez emphasized that the model's value lies in moving beyond broad screening toward precision — helping health systems direct limited resources toward individuals most likely to benefit, including those who fall outside traditional risk profiles.
The critical next step is real-world testing. Researchers plan clinical trials to determine whether patients flagged as high-risk actually engage with prevention programs — and whether that engagement reduces diabetes incidence. The distance between what a model can predict and what it can prevent in practice remains the defining question ahead.
Researchers working with electronic health records from more than three million patients have built a machine-learning model that can identify who will develop type 2 diabetes up to a decade before symptoms appear. The findings, presented this week at the American Diabetes Association's annual scientific conference in New Orleans, suggest a way to catch the disease far earlier than current screening methods allow—and to direct prevention efforts toward the people most likely to benefit.
The scale of the problem is staggering. More than 60 percent of American adults carry risk factors for type 2 diabetes, yet existing prevention programs reach only a fraction of them. The disease often develops silently over years, giving neither patients nor their doctors clear warning signs until significant damage has already occurred. Healthcare systems struggle to identify which at-risk individuals should be prioritized for intervention, meaning many people who could prevent or delay the disease never get the chance to try.
The study tracked 3.36 million adults aged 18 to 70 who received care at Kaiser Permanente Northern California between 2012 and 2024. The median age was 39, and 55 percent were women. Researchers used a statistical approach called hazard-based super learning, which combines multiple survival-analysis models to estimate each person's risk of developing diabetes over the next one, three, and ten years. The model drew on routine clinical data—age, weight, blood glucose levels, medical history, medications—along with publicly available information about neighborhood access to healthy food and walkable streets.
Over a median follow-up period of 5.4 years, the model proved remarkably accurate. When tested on new data it had never seen before, it achieved an area under the curve score of 0.883, a measure of predictive accuracy that approaches the theoretical ideal. At a threshold defining high risk—greater than 1.2 percent chance of developing diabetes within ten years—the model correctly identified 74 percent of people who would actually develop the disease while avoiding false alarms in 82 percent of those who would not. For one-year predictions, the model's estimates matched reality almost perfectly: it predicted a 1.03 percent risk, and the actual rate was 1.01 percent.
Luis Rodriguez, the study's lead author, framed the implications carefully. The model offers a way to move beyond one-size-fits-all screening toward precision identification of individuals most likely to benefit from prevention programs. It could help clinicians and health systems focus their limited resources on people who would otherwise slip through the cracks—those who don't fit traditional risk profiles but are genuinely at high danger.
The next step is clinical testing. Researchers plan to deploy the model in real healthcare settings to see whether it actually changes behavior: whether patients identified as high-risk engage more readily with prevention programs, and whether that engagement translates into fewer cases of diabetes. That gap between what a model can predict and what it can prevent in practice remains the crucial unknown. But if the model proves effective in the real world, it could reshape how healthcare systems approach one of America's most costly and preventable chronic diseases.
Notable Quotes
The model has the potential to create an opportunity for clinicians and health systems to focus prevention efforts on the high-risk individuals often missed by traditional screening— Luis A. Rodriguez, PhD, MPH, RD, lead author
The Hearth Conversation Another angle on the story
Why does a ten-year prediction window matter so much? Couldn't doctors just screen people annually?
Because diabetes develops quietly. By the time someone has obvious symptoms, their pancreas has already been struggling for years. A ten-year window gives you time to actually change someone's trajectory—diet, exercise, weight loss—before the disease takes hold. Annual screening catches people too late.
The model uses neighborhood data about food access and walkability. That seems like it's saying poverty causes diabetes.
It's more subtle than that. The model is recognizing that risk isn't just biological. Where you live shapes what you eat and how much you move. The model accounts for that reality. It's not blaming anyone; it's being honest about the conditions people live in.
Eighty-two percent specificity sounds good, but that means one in five people flagged as high-risk won't actually develop diabetes. Aren't you scaring people unnecessarily?
That's a real tension. But consider the alternative: missing the 74 percent who will develop it. And for someone flagged as high-risk, the recommendation isn't medication—it's lifestyle change that's beneficial regardless. The false alarms aren't harmless, but they're not catastrophic either.
Why does this need to be tested in clinical settings? You already know it predicts accurately.
Because prediction and prevention are different things. A model can be mathematically perfect and still fail in the real world. People might ignore the warning. Doctors might not know how to act on it. Prevention programs might be full. The model is only useful if it actually changes what happens to patients.
Three million patients is a lot, but they're all from one health system in California. Does this work everywhere?
That's the honest limitation. Kaiser Permanente Northern California is relatively affluent and well-integrated. Whether the model works in rural areas, in different racial and ethnic populations, in fragmented healthcare systems—that's unknown. The next phase will tell us.