Artificial Intelligence Versus Human Coaching in Diabetes Prevention
In this expert Q&A, Nestoras Mathioudakis, MD, MHS, associate professor of medicine at Johns Hopkins University, discusses findings from his presentation at the 2025 American Diabetes Association Scientific Sessions on the use of artificial intelligence (AI) versus human coaching in diabetes prevention. He outlines results from a randomized controlled trial comparing a fully autonomous, app-based AI intervention with a CDC-recognized human-led Diabetes Prevention Program (DPP). Dr Mathioudakis highlights that the AI intervention performed nearly identically to the human program, suggesting a promising, scalable alternative to the underutilized DPP model. He also explores the implications for clinical practice, barriers to traditional program uptake, and future directions in patient-centered AI research.
Additional Resource:
- Mathioudakis N. Artificial intelligence vs. human coaching for diabetes prevention—results from a 12-month, multicenter, pragmatic randomized controlled trial. Abstract 1956-LB. Presented at: American Diabetes Association 85th Scientific Sessions; June 20–23, 2025; Chicago, IL. https://professional.diabetes.org/scientific-sessions
TRANSCRIPTION
Nestoras Mathioudakis, MD, MHS: My name is Nes Mathioudakis. I'm an associate professor at Johns Hopkins in the Division of Endocrinology, Diabetes, and Metabolism. And I will be presenting our research on artificial intelligence versus human coaching for diabetes prevention. This is a result of a randomized controlled trial that we recently completed.
Consultant360: Can you walk us through some of the key objectives and the design of that trial?
Dr Mathioudakis: So our study is addressing—prediabetes is a highly prevalent condition. One in three US adults, or maybe even more, are affected by prediabetes. And so this is a high-risk condition that is associated with an increased risk of developing type 2 diabetes. So every year about 5 to 10% of people with prediabetes will go on to develop type 2 diabetes. And there are opportunities to reverse one's prediabetes to normal, or to slow down the progression to type 2 diabetes.
We know that lifestyle interventions modeled on the landmark Diabetes Prevention Program (DPP), which is a 12-month lifestyle change program, are highly effective in managing prediabetes and preventing the risk of progressing to diabetes. But unfortunately, the DPP is highly underutilized, so it's been estimated that fewer than 1% of eligible people with prediabetes are referred to a lifestyle change program. And there are many reasons for this. It's a big commitment. It's a 12-month program. You meet weekly for 6 months, and then every other week for the second 6 months. There's scheduling issues, transportation issues, and of course just competing life priorities. So less than 1% of people go to the DPP.
We were interested in exploring whether a fully automated AI-based diabetes prevention program, which a person could do at their own pace, could achieve results comparable to the standard-of-care DPP. And I think what was really innovative about the study was the fact that there had been previous randomized controlled trials of digital diabetes prevention programs, but none were fully autonomous, and none of them really had a design where they were comparing to the standard of care, which is an actual CDC-recognized diabetes prevention program.
So they might give patients brief counseling, like a brochure or brief education, but not the actual full evidence-based intervention. And many of the previously published results in this area, actually they present their results on program completers—people who successfully make it through the program—which kind of inflates the results. So we were interested in kind of conducting a pragmatic, real-world study where patients were referred either to a human coach–based diabetes prevention program, or a fully automated AI DPP in the form of a mobile app and a digital body weight scale.
C360: Over the 12-month trial, were there any surprising results or challenges?
Dr Mathioudakis: Yeah, so we randomized 368 adults 1:1. The study was designed to be a non-inferiority study, which was meant to show that the AI was no worse than the human. And actually the results that we achieved at 12 months were identical. So 31.7% of participants in the AI arm met the primary endpoint, and 31.9% in the human group met the endpoint.
And so the primary endpoint was achieving what the CDC has defined as the type 2 diabetes risk reduction benchmark. That's either losing 5% of your weight or losing 4% body weight and doing 150 minutes per week of physical activity or reducing your A1C by 0.2 or more from baseline. Participants could have met any of those to be considered having done sort of successful lifestyle intervention to reduce their risk in diabetes. And the results were pretty much identical. The study was not powered for demonstrating equivalence. So it was a non-inferiority design, but most likely had it been powered as such, we would have met an equivalence margin because these results were almost identical.
C360: Why do you feel now is the right time to explore AI coaching as an alternative or complement to traditional programs?
Dr Mathioudakis: Yeah, so this is obviously a very hot topic, right? AI, since the emergence of large language models and ChatGPT in, like, 2021, 2022, and thereafter, has really exploded in medicine, and the applications are everywhere in imaging disciplines and cognitive specialties. But very few AI-based interventions have actually been evaluated against the standard-of-care human-based delivery. And so, you know, this is actually one of a handful of trials to have done that.
There's huge potential for AI, obviously. You know, in terms of lifestyle change, now that we have very authentic-sounding chatbots that really can feel like a lifestyle coach and that can make it very convenient for you to make diet or physical activity changes on your own time, it's really appealing to consider this as a potential alternative to the standard, given its constraints. But it really needed to be rigorously explored, and no one had before this done that.
We're excited not just that these results showed value for prediabetes, but potentially to demonstrate a model for other types of AI-based interventions to demonstrate efficacy against a human gold standard.
C360: What were some of the most important outcomes or findings from the study that you feel clinicians should take away as they consider these kind of digital interventions in their own practice?
Dr Mathioudakis: Yeah, that's a great question. So we really tried to conduct this trial in a very pragmatic way. So our research team really did not interfere with the participants’ engagement in their assigned program after they were randomized. So it was up to them to either schedule their first class with the human DPP or to register in the app—which was developed by a company called Switch Health—and start that process. We were hands-off. Once referred, you do your thing, and let's see what happens.
And a couple things were, I think, striking about the patterns of engagement. First and foremost, it's much easier to start an AI-based DPP. So in this trial, we had patients wear ActiGraph devices throughout the study. And so we intentionally withheld giving them the app and the scale for the first week because we wanted to measure their baseline physical activity. But we could have designed the study where, you know, at randomization they were given a scale and downloaded an app and started right away.
The human DPP group—you know, we referred patients to one of four DPPs. And for these lifestyle classes, they have to get a cohort of a certain size before they can get going. And many patients faced scheduling barriers, getting into a class. So there was a delay of several weeks before they could even get in. So when we look at the time to initiation, it's much, much easier to start a digital DPP than a human DPP.
And so from a primary care physician or referring physician standpoint, there's obviously an obvious appeal to be able to say, "Hey, you've got prediabetes. I have this proven program that can help, and you can start it today. You don't need to call around and find a program and make your scheduling work."
C360: What are the most critical areas here for future research, either based on limitations of this study or questions that emerged in your findings?
Dr Mathioudakis: Yeah, excellent question. So this study was conducted at two US sites—Baltimore, Maryland at Johns Hopkins, and Reading Tower Health in Pennsylvania. We think we got a pretty diverse demographic. Reading is more of a rural-suburban population, and Baltimore is more of an urban-suburban. But first, we need larger studies in more diverse settings to make sure that this really can generalize broadly in the United States and then even outside the United States.
This was designed as a non-inferiority trial, and it would be nice to demonstrate equivalence with a larger powered study. And I think results would go that way with larger samples.
And then third, we tested one AI-based program. This was an app, as I mentioned—Switch Health. There are other AI-based interventions, and they may perform differently. They have different functionality. And so understanding, are all of them equivalent? Is it just this one? So more sort of head-to-head studies of different AI interventions.
And then finally, really understanding the role of patient preference in this type of research. So there are some people who may really need human-based intervention. They like the social interaction, the accountability of checking in with a human lifestyle coach. And there may be others who really don't want that, and they want the flexibility that the AI offers. So trying to do more research in a priori—like up front, before you assign the program—figuring out which is a better fit for a person, I think, will be an important next step in research in this space.
C360: Anything else you want to add or think is important to include in this conversation?
Dr Mathioudakis: The only other thing I would add is that one other strength of this study is that physical activity was not based on self-report, unlike in the original Diabetes Prevention Program. We actually measured this using blinded ActiGraph devices in both arms. So once per month throughout the 12-month trial, participants were asked to wear this device for seven consecutive days. And we used that physical activity to average their physical activity levels over the 12 months in evaluating that outcome of 150 minutes per week of physical activity. So I think that that's a real strength, that it was an objective metric of physical activity.