Prediabetes is a reversible metabolic condition between normoglycemia and type 2 diabetes, defined as impaired fasting glucose, impaired glucose tolerance, or elevated A1c below the diagnostic range for diabetes.

Prevention relies on structured lifestyle programs delivered by dietitians, exercise physiologists, behavioral specialists, and endocrinologists, with regular monitoring of weight and glycemic indicators.

Many people struggle to participate because of time demands and the need for sustained engagement.

A team led by Nestoras Mathioudakis, MD, MHS, Division of Endocrinology, Diabetes, and Metabolism of the Johns Hopkins University School of Medicine in Baltimore, evaluated whether an exclusively artificial intelligence (AI)-led lifestyle intervention based on the Diabetes Prevention Program (DPP) vs a human-delivered DPP to determine whether the fully automated approach was noninferior.

The study assessed whether adults with prediabetes and overweight or obesity could achieve targets for weight loss, A1c reduction, and weekly physical activity through either approach.

This phase 3, parallel group pragmatic noninferiority randomized clinical trial was conducted at two US sites and enrolled adults ≥ 18 years with prediabetes and overweight or obesity. The AI app delivered personalized notifications on weight management, physical activity, and nutrition, informed by actively collected data such as weight measurements and meal logging and passively collected data such as geolocation and accelerometry. 

Physical activity was tracked through a smartphone or wearable device, meals were logged through a food library or photo-based detection, and weight was recorded automatically via a connected scale or entered manually.

A total of 368 participants were included (median age, 58 years; 71% women; 61% White; median BMI, 32.3).

After referral, 93.4% initiated the AI-led DPP and 82.7% initiated the human-led DPP.

The primary outcome was achieved by 31.7% in the AI group vs 31.9% in the human-led group (risk difference, -0.2%; one-sided 95% CI, -8.2%), meeting the criterion for noninferiority. 

The composite endpoint required maintaining A1c < 6.5% throughout the study. Participants also had to meet one of three goals at 12 months: ≥ 5% weight loss; ≥ 4% weight loss plus ≥ 150 minutes of weekly moderate to vigorous activity; or an A1c drop of ≥ 0.2 percentage points.

A1c remained > 6.5% in 4.4% of those in the AI group compared with 3.8% of those in the human-led group.

Additionally, program completion was higher in the AI arm, with 63.9% completing the AI-led DPP compared with 50.3% completing the human-led DPP.

The authors noted that a fully automated AI-led DPP can match the effectiveness of human coaching while offering broader reach. Although the AI program did not outperform the human-led model, it delivered comparable outcomes for weight, activity, and A1c and drew higher participation, indicating that AI may expand access to structured prevention support for people who might not engage in traditional programs. Further research should clarify its impact on diabetes incidence, performance in diverse populations, cost implications, and where AI-based tools best fit alongside established DPP models.

Broader Evidence

A 2025 review reported generally positive findings for AI programs designed to slow the progression from prediabetes to diabetes, although only six of the 20 intervention studies were randomized. Programs relying solely on dietary advice without broader lifestyle components have produced inconsistent results. Evidence indicates that AI can support diabetes prevention and management, but interventions must be tailored to their intended goals.

The authors concluded that among adults with prediabetes and overweight or obesity, a fully automated AI program performed as well as a human-delivered program in supporting weight loss, increasing physical activity, and lowering A1c. Additional studies are needed to assess diabetes incidence, applicability to other AI tools, costs, use in diverse communities, and participant preferences.

This story was translated from MediQuality, part of the Medscape Professional Network.