Editorial 

AI in Medicine

Artificial Intelligence and Diabetes Prevention

Leigh Perreault

JAMA Published Online: October 27, 2025

doi: 10.1001/jama.2025.20099

As of 2021, an estimated 38.4 million people in the US had diabetes, and 97.6 million with prediabetes were at increased risk of diabetes.¹ Approximately 70% of people with prediabetes develop diabetes over their lifetime.^2,3 Furthermore, there is evidence that people with prediabetes have increased risk of the microvascular and macrovascular complications of diabetes, even if they do not meet diagnostic criteria for diabetes.^4,5 Hence, the American Diabetes Association supports screening, diagnosis, and intervention for people with prediabetes.⁶These recommendations are the direct result of findings from landmark trials such as the US Diabetes Prevention Program (DPP) that enrolled participants with prediabetes and tested strategies to prevent or delay the progression to overt diabetes. Major findings included reduction of onset of type 2 diabetes by 58% with lifestyle modification and 31% with metformin compared with placebo during the 3-year DPP randomized trial, and durability of the initial interventions to prevent or delay diabetes for more than 20 years after randomization in the extension DPP Observational Study, as well as cost-effectiveness of intensive lifestyle modification and health care cost savings from metformin.⁷

The Centers for Disease Control and Prevention (CDC) sought to translate the findings of the National Institutes of Health–funded DPP to the public. Hence, the lifestyle curriculum developed for the DPP was adapted for dissemination by the CDC-supported National DPP (NDPP) in 2012.⁸ To date, the NDPP has reached nearly 1 million individuals through approximately 1500 recognized delivery organizations in 50 US states.⁹ Despite its diversity of settings, flexibility in curriculum delivery, and eligibility for reimbursement through third party payers, 99% of individuals with prediabetes do not access the NDPP.⁹

In this issue of JAMA, Mathioudakis and colleagues investigated whether artificial intelligence (AI) could help address this implementation gap.¹⁰ In a phase 3, open-label, parallel group, noninferiority randomized clinical trial, the authors randomized adults with overweight or obesity and a hemoglobin A_1c (HbA_1c) of 5.7% to 6.4% to receive either referral to a DPP-modeled AI-led lifestyle intervention or referral to a CDC-recognized human-led DPP. The AI-led DPP involved mailing participants a digital health kit, which included an activity-tracking device and scale with instructions to link these to an app that delivered notifications for diet, exercise, and nutrition. The primary outcome was a composite surrogate end point of diabetes prevention that included maintenance of HbA_1c less than 6.5% plus weight loss of at least 5%, weight loss of at least 4% and physical activity of at least 150 minutes per week, or reduction in HbA_1c by at least 0.2 percentage points at 12 months.

Among 427 individuals screened for eligibility at 2 health systems, 368 were randomized and 313 (85%) completed the trial over 3 years.¹⁰ Trial participants had high baseline physical activity level and education, and a high percentage were female and White, limiting generalizability. Results demonstrated that the fully automated AI-led DPP was noninferior to a human-led DPP, with 32% of participants achieving the composite outcome (58 of 183 participants [31.7%] in the AI-led DPP group and 59 of 185 [31.9%] in the human-led DPP group; risk difference, −0.2% [1-sided 95% CI, −8.2%]). The 3 individual components of the composite outcome were directionally consistent with the results of the primary outcome. Given that the AI-led DPP was not more effective than the human-led DPP, the authors conclude that the value of the AI-led DPP lies in its scalability, rather than its superiority.

The study design could also be described as an effectiveness-implementation hybrid type I randomized trial, in which case metrics of implementation science can be applied to speculate on its value if scaled. Implementation science is the scientific study of methods to promote the adoption and integration of evidence-based practices into routine care and public health to improve population health. Implementation studies are evaluated based on their ability to meet the metrics of RE-AIM (reach, effectiveness, adoption, implementation, and maintenance/sustainment).¹¹

For AI-led DPP to be scaled, it would need to reach people at risk for diabetes, be effective at preventing diabetes, be adopted by people at risk, be implemented in a way that is accessible and acceptable to people at risk, and be maintained long term. The work by Mathioudakis and colleagues does indeed meet most of these metrics—all but reach.

Effectiveness was declared a priori as the primary outcome for this study. Nevertheless, effectiveness was heavily influenced by adoption and maintenance rates. Initiation (adoption) was higher in the AI-led DPP group than in the human-led DPP group (93.4% vs 82.7%), as was study completion (maintenance; 63.9% vs 50.3%). Accordingly, it can be gleaned from the data that the human-led DPP was more useful but used less often, whereas the AI-led DPP was less useful but used more often, ultimately rendering similar effectiveness. The degree of effectiveness, at 32% of participants meeting the composite end point, is clinically meaningful because this approximates program goals of CDC-sanctioned community-based in-person or distance curriculum delivery¹² without the need for human coaches.

Implementation of the AI-led DPP appears to have been successfully conducted with fidelity. Perhaps serendipitously, the human-led DPP was adapted to distance learning during the COVID-19 pandemic through virtual sessions. Digital literacy improvements and acceptability of AI likely enhanced referral uptake of both groups. Interestingly, the human-led DPP group saw higher-than-typical participation rates when delivered virtually compared with historical in-person rates. Hence, the AI-led intervention may have demonstrated superiority in effectiveness if the human-led DPP had been delivered in person as originally planned. Noninferiority using AI suggests acceptability of this technology and may set a precedent for future implementation work.

Implementation science strives to shorten the time it takes evidence-based medicine to reach clinical practice.¹¹However, most implementation studies overestimate the power of the intervention and the strength of the implementation strategies designed to support benefit from the intervention.^13,14 The study by Mathioudakis et al successfully achieved its goals and contributes important new findings about how an AI-delivered program that is as effective as the current standard of care, better adopted and maintained, and potentially scalable at low cost can be used to prevent diabetes.