Diabetes care: Are automated AI interventions as effective as human coaching?

A fully automated artificial intelligence (AI)-driven diabetes prevention program achieved an outcome comparable to human coaching in a phase 3 trial of 368 adults with prediabetes who were overweight or obese. 

At 12 months, 31.7% of patients referred to the artificial intelligence (AI) program and 31.9% referred to a human-led diabetes prevention program achieved the composite outcome of maintaining hemoglobin A1c levels below 6.5% and met at least one of the following: at least 5.0% weight loss; at least 4.0% weight loss plus at least 150 minutes of weekly physical activity; or an absolute reduction in hemoglobin A1c of at least 0.2%.

Results appeared consistent across each component of the composite end point and in sensitivity analyses. The findings, published in JAMA by Nestoras Mathioudakis, MD, of Johns Hopkins University School of Medicine in Baltimore, and colleagues, met the prespecified criterion for noninferiority, suggesting that “a fully automated AI-led [diabetes prevention program] may be a viable alternative to a [diabetes prevention program] led by human coaches.”

Participants were randomly assigned to referral to either an AI-led program delivered through a mobile app with a Bluetooth-enabled scale or a Centers for Disease Control and Prevention-recognized human coach-led diabetes prevention program delivered remotely over 12 months.

The AI platform used a reinforcement learning algorithm—learning which prompts, timing, and content elicited greater user engagement—to personalize push notifications for weight management, physical activity, and nutrition based on both actively (eg, weight measurements, meal logging) and passively (eg, geolocation, accelerometry) collected data.  

Engagement Differences

Although outcome achievement rates were similar between the groups, engagement levels differed. Program initiation occurred in 93.4% of patients referred to the AI program vs 82.7% referred to the human-led program, a statistically significant difference. The proportion of participants meeting program completion criteria was also significantly higher with the automated intervention (63.9% vs 50.3%).

The researchers suggested the asynchronous format may reduce barriers such as travel, scheduling constraints, and limited cohort availability.

 In an interview with AACE Endocrine AI, Dr. Mathioudakis discussed how clinicians might choose between the two formats.

“We don’t yet have a validated triage tool to determine who would prefer an [AI-led] or traditional human-led program, and in our trial neither age nor baseline tech comfort reliably predicted who did better with either modality,” he said. “In practice, it’s a shared decision driven by logistics and preference. If a patient needs flexibility because of work or caregiving, an AI-based, asynchronous program may be a better fit; if they value live interaction and peer accountability, a coach-led group can be ideal.”

Commentary Raises Implementation Questions

In a related commentary, Esteban Zavaleta-Monestel, PharmD, and Sebastián Arguedas-Chacón, PharmD, both of Hospital Clínica Bíblica, San José, Costa Rica, wrote that the study “represents an important step toward understanding how AI can enhance behavioral interventions for chronic disease prevention and shifts perspective from viewing AI as a supplementary tool to recognizing it as an autonomous participant in preventive care.” However, the commentators cautioned that the trial’s noninferiority margin of 15 percentage points “could represent a clinically meaningful difference when projected onto long-term diabetes incidence.” They also suggested the composite behavioral outcome could have masked differences in clinically relevant outcomes.

Dr Mathioudakis offered a similar perspective, telling AACE Endocrine AI: “Trial efficacy doesn’t guarantee real-world impact.” He added that indicators would include broader and more equitable reach, with more people with prediabetes enrolling across geographic regions and among racial, ethnic, and socioeconomic groups historically underrepresented in diabetes prevention programs; sustained engagement and achievement of validated risk-reduction targets such as 5% to 7% weight loss, increased physical activity, and improved hemoglobin A1c; and hard outcomes and population-level effects over 2 to 5 years, including lower progression to type 2 diabetes, comparable outcomes across subgroups, and demonstrable cost-effectiveness.

“The most convincing signal would be population-level declines in new-onset diabetes in communities where [AI-led diabetes prevention programs] are widely implemented,” Dr Mathioudakis said.

The commentary authors also cautioned that the trial population’s baseline digital literacy and access to mobile technology may limit generalizability to lower-resource settings or older populations. They wrote that “ensuring accessibility and inclusion is essential to prevent the widening of existing health disparities.”

This will require equity to be built in from the start, Dr. Mathioudakis noted. “Programs should offer plain-language, multilingual content that adapts to different literacy levels and can be accessed through multiple channels—SMS, phone, web, and offline-capable apps—so participation doesn’t hinge on owning a smartphone or broadband,” he said.

He added that health systems can provide onboarding and ongoing technical support through digital navigators or community health workers. They can also partner with payers to cover devices or data plans and create community access points in clinics and libraries. Tracking enrollment, engagement, and outcomes across demographic groups will also be important, as will iterating on AI-led programs with community input to ensure they are appropriate for people with different levels of health literacy and from diverse cultural backgrounds.

The commentary authors also questioned whether automated systems can fully replace human coaching. “Behavioral modification depends on interpersonal reinforcement and contextual understanding, capacities that current AI systems only partially replicate,” they wrote, concluding, “Automation should complement, rather than replace, the human connection that remains fundamental to effective health promotion.”

For full disclosures of the researchers, visit jamanetwork.com. The commentary authors reported no conflicts of interest.

AACE Endocrine AI is published by Conexiant under a license arrangement with the American Association of Clinical Endocrinology, Inc. (AACE^®). The ideas and opinions expressed in AACE Endocrine AI do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.