An artificial intelligence (AI)-led Diabetes Prevention Program (DPP) was as effective as a traditional human-led program in achieving recommended goals for weight loss, A1c reduction, and physical activity, according to a randomized trial of adults with prediabetes and overweight or obesity.

Prediabetes is pervasive in the US, but access to evidence-based lifestyle interventions is limited, said lead study author Nestoras Mathioudakis, MD, associate professor of medicine at Johns Hopkins University, Baltimore.

The study, published in JAMA, compared referrals to an AI-powered DPP lifestyle intervention with referrals to a human-coached DPP. The AI intervention was delivered via a mobile app and Bluetooth-enabled digital scale and provided personalized push notifications based on user-supplied data on weight, diet, and activity levels.

One example of a push notification: “Looks like you’re at the grocery store, Rita! Want a quick list of high-fiber snacks or smart swaps to stay on track this week?”

The app also provided location- and goal-based education, with gamification elements to promote engagement.

AI and Human Coaching Produce Similar Results

Researchers included 368 participants, who had a median age of 58 years and median BMI of 32.3 (72.3% had obesity and 27.7% had overweight). Of the participants, 71% were women, 61% were White individuals, 27% were Black individuals, and 6% were Hispanic individuals.

The study’s primary outcome was a composite of several diabetes risk factors, including maintaining an A1c < 6.5% through the study and achieving ≥ 5% weight loss, at least ≥ 4% weight loss plus at least 150 minutes of physical activity each week, or a reduction in absolute A1c of ≥ 0.2% after 12 months.

Approximately one third of participants in both the AI and human-led groups achieved the primary outcome (31.7% and 31.9%, respectively). Results were consistent across sensitivity analyses and individual components of the composite endpoint.

“The proportions achieving the composite endpoint — and each of its components — were nearly identical, and the AI-led program far exceeded our prespecified noninferiority margin,” Mathioudakis told Medscape Medical News.

Applying AI to Expand Access to Care

“Asynchronous, app-based delivery offers a way to scale the DPP by allowing individuals to engage at their own pace,” said Mathioudakis. He added that, despite the growing enthusiasm for AI in medicine, head-to-head comparisons between this technology and established human-delivered programs remain limited.

The study began in October 2021 and concluded in December 2024. Because of the COVID-19 pandemic, all human-led DPP sessions were conducted via videoconference. This shift, Mathioudakis said, may have narrowed differences between groups.

“Even so, outcomes with the asynchronously delivered AI program were not inferior to the synchronous remote, human-led DPP,” he said.

Participants also displayed distinct engagement patterns. Compared with human-led DPP, more individuals initiated the AI-led program (82.7% vs 93.4%; P = .001) and met engagement criteria for program completion (50.3% vs 63.9%; P = .008).

For Mathioudakis and his colleagues, this suggested that AI-led DPP was potentially a more convenient option for most middle-aged and older adults.

Although the findings are promising, expanding the use of AI-led DPPs would require policy changes from the CDC, which currently require some degree of human coaching in recognized DPPs.

“Given that our trial showed comparable outcomes to human-led delivery and results consistent with real-world DPP performance, we believe this policy should be revisited,” he said.

The AI-led DPP used in this study, Sweetch Health, is not yet a CDC-recognized program. Other AI-based DPPs, such as Lark, are recognized but still incorporate human coaching in addition to the AI component, Mathioudakis said. The CDC would need to waive the requirement that programs offer human coaching in order for more fully automated AI-DPPs to be reimbursable.

Mathioudakis added that the study population, drawn from two East Coast sites, was highly motivated and had relatively high baseline physical activity, which may limit generalizability. Larger, more diverse studies are needed to confirm the findings.

“As more AI-based programs emerge, head-to-head comparisons among different AI-DPPs will be informative. An AI-led approach will not suit everyone; some individuals benefit more from human interaction and accountability,” said Mathioudakis, adding that future research should focus on best matching patients to the modalities they prefer.

Ensuring Effectiveness, Overcoming Bias

“AI interventions can be very impactful and provide human-like interactions, and other studies have shown them to be effective in other aspects of healthcare,” said Kevin Pantalone, DO, director of diabetes initiatives at the Cleveland Clinic, Cleveland who was not involved in the study. “However, we cannot simply enact AI-based interventions in healthcare by simply assuming noninferior outcomes vs the current standard of care/usual care approach; research is necessary to make that determination.”

Pantalone noted that many patients are already comfortable with technology and regularly leverage AI tools in their daily lives, which could facilitate adoptions of AI-DPPs.

However, healthcare provider bias remains a major barrier.

“Healthcare providers sometimes assume that a technology-based intervention may not be effective or may be too complicated for some individual patients, so the intervention is not offered,” Pantalone told Medscape Medical News. “The use of AI-based interventions should be reviewed with all eligible patients, and a shared decision-making process enacted to determine what is best for any one individual.”

This study was funded by the National Institute of Diabetes and Digestive and Kidney Diseases and the National Institute on Aging. Additional support came from the Johns Hopkins Institute for Clinical and Translational Research, partially funded by the National Center for Advancing Translational Sciences.

Mathioudakis disclosed no financial conflicts of interest.

Pantalone disclosed no financial conflicts of interest.