Get In-depth Biotech Coverage with Timmerman Report.
30
Oct
2025
The Outsized Significance of A New Study of AI in Diabetes Prevention

David Shaywitz
A lifestyle intervention delivered by AI was found to be as effective as validated traditional interventions delivered by trained experts in a carefully conducted implementation study conducted by Johns Hopkins researchers and just published in JAMA.
These results have broad significance and speak to the promise of AI to deliver promising behavior-change interventions at unprecedented scale.
Context: The DPP
The core of the study involved a striking and unexpected finding from a classic study, the Diabetes Prevention Program (DPP), conducted two decades ago, and led by Dr. David Nathan of the MGH Diabetes Center.

Dr. David M. Nathan
The DPP was designed to examine whether one of three potential interventions administered to patients at high risk for diabetes could prevent or delayed their progression to the full-blown disease, compared to placebo.
The interventions examined by the DPP originally included one of two medicines (metformin or troglitazone), although the troglitazone was soon dropped due to safety concerns. The third intervention evaluated was intensive lifestyle coaching.
In 2002, the DPP investigators reported that intensive lifestyle coaching reduced incidence of diabetes by a staggering 58% compared to placebo over three years; metformin reduced incidence by 31% compared to placebo over the same time.
This 58% reduction has become the benchmark – a consistently elusive benchmark – that researchers and companies have chased ever since.
As I discussed in Forbes in 2012, in piece about the (at the time, nascent) digital health company, Omada Health leveraging digital tools to try to emulate the DPP success, the secret sauce of the DPP is the intensity and dedication of the coaching provided.
I had particular visibility into the nature of this coaching since my endocrinology training was at MGH, where I was mentored in diabetes by Dr. Nathan and his colleagues at the MGH Diabetes Center, and learned about the nature of the DPP lifestyle interventions directly from some of the coaches who administered them.
My takeaway: these coaches were determined and relentless, deeply dedicated to connecting with their patients individually, understanding their challenges, and urging them onwards. From what I saw, there was nothing pro-forma or box-checking about any of it – the coaches (compassionately) pestered patients, exhorted them, humored them, listened to them, and believed in them – ultimately to great effect.
But duplicating these efforts – and this commitment – has been challenging.
Even when the same protocol has been followed, and coaches appropriately trained, the results have never been quite as impressive, and usually come in at about half the efficacy seen in the original DPP. This impact is still clinically meaningful, to be sure, but not at the level seen in the DPP.
Challenges include matching the original intensity (particularly when trying to deliver the program with less resources to larger populations) and maintaining participant engagement and retention.
AI- vs Human-Delivered DPP
The latest study randomized patients at high risk of diabetes into two groups: one received a DPP-style intervention administered by AI, while patients in the other group received an DPP-protocol intensive intervention delivered by human coaches. The initial intent was for these sessions to be in person, but this was changed to virtual sessions due to COVID, a change that may have actually helped the study, since as an excellent editorial by Dr. Leigh Perreault accompanying the JAMA publication observed, “the human-led DPP group saw higher-than-typical participation rates when delivered virtually compared with historical in-person rates.”
As the study authors note, “The AI used in the intervention consisted of a reinforcement learning algorithm that did not use large language models. It personalized messaging by continuously learning which prompts, timing, and content elicited greater user engagement.”
The result of the non-inferiority study, using a composite endpoint incorporating measures of glucose control, physical activity, and weight loss, was a tie: the two approaches delivered similar outcomes: about 32% of patients in each group achieved the pre-specified outcome, suggesting that for this population, the AI intervention did as well as the human coaches.
A closer look at the results highlights an important nuance: fewer patients randomized to the human coaches actually initiated the program (only about 83%), while more than 93% of patients allocated to AI coach started that intervention. Doing the math, this means that the human coaches were slightly more effective on the patients they actually reached – but the AI reached more people, and these effects netted out to the same impact.
Why This Matters
The ability of AI to achieve the same efficacy for a study population as human coaches using a well-validated approach has significance for at least three reasons that take us far beyond an encouraging proof of principle.
- Scale: AI was built to scale; the cost for delivering an AI intervention to ten patients, 100 patients, or 10,000 patients almost certainly will be far less expensive than human coaches would be, coaches who must be individually trained, organized, supported, etc. The AI will also be far more consistent in its content delivery.
- Constant learning and improvement: an AI system also will facilitate the rapid, continuous learning upon which medical progress often depends. Not only can new best practices be identified, but they can be incorporated immediately at scale. It is also much easier to set up comparisons between slightly different approaches, enabling ongoing refinement. Thus, whatever level of performance the AI achieves at the start is likely to improve significantly over time.
- Supercharging promising local approaches: it is not uncommon in medicine (or in any other pursuit) for a motivated provider, craftsperson, or inventor somewhere to figure out a better way of doing something, such as caring for patients with a particular type of condition. Their solution might be highly effective in the provider’s hands, but hard to transmit outside their local practice. The promise of AI-delivered interventions is to enable these insights to be systematized (perhaps in partnerships with a motivated startup), deployed as a pilot, evaluated, iteratively improved, and if effective, rapidly scaled, potentially benefiting far more patients.
An important question I continue to have is whether this approach will plateau at a point that is still below the level of the most expert practitioners who deliver the intervention in person.
I think about the next-level DPP coaches at the MGH Diabetes Center, for example, or the late, next level cystic fibrosis doctor Warren Warwick at Fairview-University Children’s Hospital, in Minneapolis, so memorably profiled in The New Yorker by Atul Gawande.
In theory, it might be possible to capture and systematize much of what the MGH diabetes coaches and Dr. Warwick have done so well, and to use technology to deliver at scale, and iteratively improve the approach to a level that matches or even exceeds that of these exemplars.
I’m curious about this possibility, but skeptical.
Gawande writes about other medical centers trying to replicate Warwick’s CF success, and reflects,
“Yet you have to wonder whether it is possible to replicate people like Warwick, with their intense drive and constant experimenting. In the two years since the Cystic Fibrosis Foundation began bringing together centers willing to share their data, certain patterns have begun to emerge, according to Bruce Marshall, the head of quality improvement for the foundation. All the centers appear to have made significant progress. None, however, have progressed more than centers like Fairview.”
The ability of technology to scale means that reasonably effective interventions delivered consistently can benefit large swaths of patients.
But it feels comforting to know, or at least to believe, that there may be something special, powerful, ineffable and deeply human that lies beyond the reach of even the most impressive digital technologies.



































