An AI-Led Diabetes Program Met the Trial’s Human-Coaching Benchmark

A fully automated lifestyle program helped adults with prediabetes reach a 12-month health benchmark at almost exactly the same rate as participants referred to human coaches.

In a phase 3 randomized clinical trial, 58 of 183 adults assigned to the AI-led Diabetes Prevention Program reached the study’s composite outcome, compared with 59 of 185 adults assigned to one of four human-coached programs: 31.7% versus 31.9%. The risk difference was −0.2 percentage points, and the result met the study’s prespecified noninferiority rule.

That is a meaningful result—but it is not proof that an app prevented diabetes, matched every benefit of a skilled coach, or will work for everyone. It shows that, for this selected group of adults, referral to one fully automated program produced a similar rate of reaching the trial’s combined weight, activity, or blood-sugar target.

Research references: [1]

The trial compared two real programs for a year

The trial ran from October 11, 2021, through December 16, 2024, at Johns Hopkins Hospital in Baltimore and Reading Hospital Tower Health in Pennsylvania. It enrolled 368 adults with laboratory-confirmed prediabetes and overweight or obesity and randomized them 1 to 1 to referral to either an AI-led or human-led program.

The median participant was 58 years old; 71% were women, 27% were Black, 6% were Hispanic, and the median body mass index was 32.3. Participants were English-speaking smartphone users, and baseline activity was relatively high: median device-measured moderate-to-vigorous physical activity exceeded 230 minutes per week in both groups.

The AI group received a commercial Sweetch Health program, a Bluetooth scale, and a mobile app. The system used weight, meal, activity, geolocation, and accelerometry data. Its reinforcement-learning algorithm—not a large language model—personalized the timing and content of prompts according to engagement.

The comparison group was referred to one of four CDC-recognized Diabetes Prevention Programs. Coaches delivered the PreventT2 curriculum remotely in scheduled group sessions. The study team initiated referrals but did not deliver either intervention or add engagement support, making this a pragmatic comparison of what happened after referral.

Research references: [1]

What “noninferior” meant here

The primary outcome required participants to keep hemoglobin A1c below 6.5% throughout the study and, at 12 months, achieve at least one of three milestones:

• at least 5% weight loss;
• at least 4% weight loss plus at least 150 minutes of moderate-to-vigorous physical activity per week; or
• an absolute A1c reduction of at least 0.2 percentage points.

Research references: [1]

How the comparison passed

The lower bound of the one-sided 95% confidence interval for the risk difference was −8.2 percentage points. Because it stayed above the prespecified −15-point margin, the AI program met the statistical definition of noninferiority.

A noninferiority trial does not prove that two interventions are identical. It asks whether the new option is worse by more than a margin chosen in advance. The study’s margin was wide enough to permit a clinically important difference, even though the observed rates were nearly identical. The outcome was also a composite of risk-related measures, not long-term diabetes incidence.

Research references: [1]

Automation made the front door easier to enter

The clearest advantage appeared before the health outcome. After referral, 93.4% of adults assigned to AI initiated the program, compared with 82.7% assigned to human coaching. More people in the AI group also met the study’s program-specific completion definition: 63.9% versus 50.3%.

Those completion figures need careful interpretation. AI completion meant engagement during at least eight weeks in the first six months and a span of at least nine months between first and last engagement. Human-program completion meant attending at least eight sessions in the first six months across a span of at least nine months. They are different forms of participation, not equal doses of support.

A companion analysis of the same randomized cohort found that participants assigned to AI began sooner—a median of 11 days after referral versus 26 days for human coaching—and engaged more evenly across the year.

That is the strongest near-term promise: structured support that can begin asynchronously, without waiting for the next group session.

Research references: [1] [2]

People still valued the human experience

Faster access did not erase the value of relationship. In the companion analysis, acceptability ratings were higher for human coaching across every reported domain, with the largest gap in satisfaction.

Preference responses also leaned toward people. Among AI-assigned participants, 46.5% said they would have preferred a human coach. Among those assigned to human coaching, 31.5% said they would have preferred a fully automated program.

Access and experience are different dimensions of care. Automated support may be easier to start and use on a person’s own schedule. Human coaching may offer accountability, interpretation, empathy, and social connection. The trial did not test a hybrid model, but its results give health systems a reason to test one.

Research references: [2]

Important boundaries

The study was unmasked and conducted at only two US sites. Participants were motivated volunteers with smartphones, English proficiency, relatively high educational attainment, and high baseline activity. Results may not transfer to people with less digital access, other languages, lower health literacy, or more complex medical needs.

The human programs shifted from planned in-person delivery to synchronous remote delivery during the COVID-19 pandemic. Sweetch Health could update app features and its interface during the trial, although the reinforcement-learning algorithm behind push notifications stayed unchanged. More adverse events were reported in the AI group, but investigators attributed none to either intervention.

The study was fully funded by the National Institute of Diabetes and Digestive and Kidney Diseases, which covered participant costs for both program types. Sweetch Health and the participating human programs were compensated for delivering their interventions. The paper states that they did not analyze the data, interpret the results, or access the full cohort’s primary-outcome results.

The trial did not establish cost-effectiveness, long-term diabetes prevention, or the right rules for privacy, consent, data retention, and secondary use in future deployments. Its registration and protocol show that the comparison was designed before the final results, while the published report remains the source for the completed trial’s exact analysis.

Research references: [1] [3] [4]

More support without forcing one model on everyone

The breakthrough is not an app that makes coaches obsolete. It is evidence that a fully automated program can clear a meaningful one-year comparison against established human-led programs for a defined group of adults—and can start people sooner.

A more human health system could use automation to widen the front door, preserve coaching as a meaningful choice, and offer human support when a participant asks for it or when self-directed engagement falters.

Future trials should test that combination directly. They should measure diabetes incidence, costs, privacy outcomes, accessibility across languages and devices, and whether the benefit holds in populations digital health programs too often miss.

*This article reports research findings and is not individualized medical advice.*