Machine-learning model identifies adults at highest risk for type 2 diabetes up to 10 years in advance, study data show
Presenter: Luis A. Rodriguez, PhD, MPH, RD, Kaiser Permanente Division of Research, Northern California
Machine-learning modeling for T2DM prediction in over 3 million adults. Abstract 2321-P. Presented June 6, 2026.
An electronic health record-based machine-learning model identified adults at the highest risk of developing type 2 diabetes up to 10 years before onset, according to a study presented at the American Diabetes Association (ADA) 2026 Scientific Sessions.
Developed and tested in more than 3 million adults, the model strongly distinguished those who would develop the disease from those who would not, achieving an area under the curve of 0.886 in the training set, according to the authors. Performance held across 1-, 3-, and 10-year prediction windows.
“These findings represent a potential advancement over existing approaches for identifying individuals at risk of developing type 2 diabetes by enabling earlier, more precise detection and supporting a more targeted, proactive approach to prevention,” said lead author Luis A. Rodriguez, PhD, MPH, RD.
More than 60% of US adults have risk factors for type 2 diabetes, far more than current prevention programs can realistically serve, according to the authors. Because the disease often develops gradually over many years without clear warning signs, health systems struggle to identify those most likely to benefit from prevention programs or early treatment.
Several diabetes prediction models already exist, but most rely on data not collected during routine care and therefore unavailable in the electronic health record, Dr. Rodriguez noted.
“This work fills this gap,” he said of the results.
The retrospective cohort study included 3,365,464 adults age 18 to 70 years receiving care at Kaiser Permanente Northern California from 2012 to 2024. Patients were followed until they developed type 2 diabetes, died, left the health plan, or reached the end of 2024.
The cohort was randomly split in a 70:30 ratio into training and validation sets, and the investigators applied a hazard-based super learning approach to estimate each patient's 1-, 3-, and 10-year risk. Predictors spanned demographics, clinical measures, lifestyle factors, comorbidities, prescriptions, and health care utilization, along with newer inputs such as metabolic dysfunction-associated steatotic liver disease and neighborhood-level measures of socioeconomic status, walkability, and food environment.
During a median follow-up of 5.4 years, type 2 diabetes incidence was 10.7 per 1,000 person-years, reported data show. The median age was 39 years (interquartile range 28–53), and 55% of patients were female.
For 1-year prediction, the model achieved an area under the curve of 0.886 (95% confidence interval 0.883–0.888) in training and 0.883 (95% confidence interval, 0.880–0.886) in validation, with near-ideal calibration (mean predicted risk 1.03% vs observed risk 1.01%, slope 1.26), the investigators said.
At the optimal cut-point of greater than 1.2% risk, which flagged the top two risk deciles, sensitivity was 80%, specificity was 81%, and the number needed to evaluate was 25. Results were consistent at 3 and 10 years, according to the authors.
“Our 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 who have the most to gain from prevention and treatment,” Dr. Rodriguez said in an ADA statement.
The model can support clinicians in identifying patients for prevention programs and pharmacologic treatment and enable efficient recruitment for intervention studies, according to the authors.
The investigators intend to test it in a clinical setting to determine whether it increases engagement in type 2 diabetes prevention programs and reduces diabetes incidence.
The team is now finalizing a temporal validation using more recent data. “We'll have those results in the coming weeks,” Dr. Rodriguez said in his presentation.
Disclosures
Dr. Rodriguez reported no relevant disclosures. Several coauthors reported research support from industry, including Freenome, Bayer AG, AstraZeneca, and Gilead Sciences.
References
Rodriguez LA, Yassin MM, Neugebauer R, et al. Machine-learning modeling for T2DM prediction in over 3 million adults. Abstract 2321-P. Diabetes 2026; 75(suppl 1). Presented at the ADA 2026 Scientific Sessions, June 6, 2026.
American Diabetes Association. Machine learning model accurately predicts long-term risk of type 2 diabetes. Press release. June 5, 2026.