A new kind of coach: How artificial intelligence is helping people prevent diabetes

When you first get the diagnosis that your blood sugar is too high, you feel as if you are standing on the edge of a cliff. The diagnosis of prediabetes means you are not sick yet, but you’re close. You may tell yourself you are going to eat better or move more, but it seems overwhelming to do that by yourself.

A newly published study by Johns Hopkins Medicine has now offered some ray of hope: an artificial intelligence–powered program can help people achieve these life changes as well or better than other traditional, human-led programs.

The study was published in JAMA, funded by the National Institutes of Health, and it is the first phase III randomized controlled clinical trial to investigate and compare an AI Diabetes Prevention Program (AI-DPP) with a human-led Diabetes Prevention Program. The goal was simple but powerful: could technology meet people where they are—in their homes, on their smartphones, any time of the day— and help them to make impactful changes?

Currently, around 97.6 million American adults are estimated to have prediabetes. If they do not make a change, many will go on to develop type 2 diabetes in the following five years. Comprehensive lifestyle programs that emphasize diet, exercise, and weight loss will reduce that risk by over 50%. The problem is that in general, so few enroll in these programs.

Barriers such as cost, time, transportation, and scheduling present real obstacles. The research team was interested in finding out if artificial intelligence could assist in overcoming some of these hurdles.

A total of 368 adults with prediabetes participated in the trial between October 2021 and December 2024. The median age of study participants was 58. Most volunteers were women, and they represented a range of races and ethnicities. Each of the volunteers was randomly assigned to one of two groups.

One group received the yearlong, standard program with a human coach through the CDC-recognized program, and the other group received an AI app with a Bluetooth scale. Each day, the AI group participants received personalized feedback relying on a “reinforcement learning” algorithm, or machine learning, that changed the app’s messages based on an individual’s behavior patterns and progress.

The app provided nudges toward small daily actions that mattered: reminders to weigh themselves, take a walk, or gently encourage participants when motivation waned. Every app interaction provided users with relevant, timely, and supportive feedback.

The participants in the human-coach group joined one of four yearlong programs that started with weekly virtual meetings and transitioned to maintenance meetings. The human coach led discussions with the participants about food, physical activity, and strategies for addressing setbacks. The yearlong programs followed the CDC’s previously researched yearlong Diabetes Prevention Program and have reported a 58% decrease in diabetes risk.

At 12 months, both subjects showed virtually identical results. Approximately 32% of subjects in both groups achieved the CDC’s composite measure for met diabetes prevention, which is defined as 5% body weight loss, 150 minutes of moderate physical activity per week, or 0.2% A1C loss.

Enrollment numbers proved even more impressive; 93% of AI participants began their program versus 83% of human coach participants. Additionally, 64 percent of those who used an AI program finished the full year, while approximately half of those in the human-led programs did so.

Lead researcher Dr. Nestoras Mathioudakis, the co-medical director of the Johns Hopkins Diabetes Prevention and Education Program, explained why he believes this is a major turning point. “Even beyond diabetes precision, there are very few randomized trials that compare AI- and patient-directed interventions, to traditional human standards of care,” he stated.

Co-first author Benjamin Lalani, now a student at Harvard Medical School, the ultimate result speaks to the importance of accessibility. “The greatest barrier to completion on the DPP is often just getting started. AI-DPPs cut out the scheduling and logistical hurdles that trip people up before they even start.

The emotional part of this story is the experience and feelings that can’t be captured in numbers: the relief, hope, and self-confidence people experienced as they incorporated the app into their everyday lives.

For someone who is trying to juggle work, family, and health concerns, the AI has flexibility that is worth countless dollars. There were no class times to remember, and there were no instructors to schedule and meet; in each instance, as long as the individual had time to focus on their health, the intervention was available to them.

The AI program’s platform utilized behavioral data collected via phones and scales to provide each person with enhanced support, customization that improved with each individual’s behavior. If the system noted someone hadn’t input their weight in a few days, it would gently remind them that they were behind. If the individual’s step count decreased to a point out of range of their average, it would encourage an individual to go for a walk in a nearby area. This perpetual availability appeared to inspire more engagement in the program than a traditional program would have inspired.

Nonetheless, programs led by humans offered something AI cannot replicate: a sense of belonging, shared experience, and a level of accountability through community. For some, simply connecting with a coach or others can be a motivating factor to remain engaged. The study did not suggest that one approach is better—it only established that both were equally effective. The main variation is in the means of connecting an individual with change.

It should be stressed that the study only included a motivated sample of tech-savvy adults who were comfortable using digital tools, and any conclusions about how the app would perform with older, younger, or less digitally literate populations are unknown. Furthermore, the trial did not assess the ultimate number of participants who developed diabetes, which would require a lengthier follow-up.

That being said, the implications are promising. Fully automated programs may help relieve a burdened health system where staffing is a challenge and reach millions of people who never show up in person to a human-led educational or behavior change experience. These digital health tools may become a feasible referral option for doctors with patients who are struggling with weight management or blood sugar management.

The team at Johns Hopkins is now exploring ongoing investigations of real-world examples of AI-based interventions. They are also examining differences in cost, engagement, and patient preferences in AI and human-centric programs. Dr. Mathioudakis captured it well, suggesting that the future of medicine may be the sweet spot between human empathy and machine intelligence.

This research suggests a potential model of diabetes prevention that is personal and scalable. AI programs can offer affordable, 24/7 support to the millions of people who do not enjoy a traditional approach to care.

Digital health tools can help improve the health fairness gap by decreasing cost and access barriers (scheduling, travel, and so forth), making health prevention efforts more equitable and accessible. For the end-user, they have the potential to provide flexibility and motivation; for a health system, they may help supplement the limitations of staff available.

With improved tools, they could help stem the growing tide of type 2 diabetes and allow people to take agency over their health—one small push at a time.

Research findings are available online in the journal JAMA.