AN AI-DRIVEN diabetes prevention programme which has faster, more consistent patient engagement than a human-coach led programme was unsatisfying for patients, according to new research.
A new randomised trial found that found that AI-powered lifestyle intervention for patients with prediabetes matched human coaching on clinical outcomes. However, it fell markedly short on acceptability and patient preference.
Diabetes Prevention Programmes (DPPs) delivered by human coaches are proven to reduce diabetes risk but remain underused because of cost and scalability constraints.
AI-driven interventions have been proposed as a way to expand access to this kind of support. However, most existing evidence on patient attitudes toward AI in medicine draws on hypothetical scenarios rather than direct, head-to-head comparisons with real-world alternatives.
Researchers at Johns Hopkins had previously found that an AI-led diabetes prevention programme was clinically non-inferior to a human-led programme, producing clinically equal results based of three factors: weight loss, physical activity, and A1C reduction.1
However, it was unclear from this research whether, despite the positive results, patients preferred this method of intervention. In this follow up study, they looked at how patients experienced each programme.
The study involved randomising 368 adults with prediabetes and obesity (or were overweight) to receive coaching through either a fully automated AI programme, delivered via the Sweetch app, or through referral to an officially recognised, coach-led DPP delivered by videoconference.2
Participants in the AI group began the intervention in an average of 11 days, markedly sooner than the human group, whose median start time was 26 days. They also showed more even, sustained engagement across the 12 months, with more participants in the AI group reaching maximum engagement.
The experience of the participants was markedly different. On a validated 30-item scale, the human-led programme scored significantly higher across every domain assessed, with the largest gap seen in overall satisfaction.
Notably, the highest acceptability score recorded for the AI programme fell below the lowest acceptability score recorded for human coaching.
Preference followed the same pattern. At 12 months, 46.5% of participants in the AI group said they would have preferred a human coach, compared with 31.5% of the human-coaching group who said they would have preferred the automated alternative.
The researchers concluded that the accessibility and asynchronous access to the AI programme which allowed users to input food data and check in freely and around their own schedules, likely drove the AI programme’s engagement advantage.
However, the human-led programme’s live, group-based format which involved open discussion with others and encouraged social interaction better supported the relational and motivational elements that participants seemed to value highly.
In comparison, the AI intervention did not attempt to simulate conversation or replicate a coaching relationship, which may partly explain the gap in preference, and suggest that future systems could narrow this divide by incorporating stronger elements of social accountability and interaction into AI models.
Overall, the study adds real-world, longitudinal evidence to a growing literature suggesting that while AI tools can extend access to lifestyle interventions to a wider number of patients, the importance of human connection and the value patients continue to place on these inter-personal relationships in behaviour-change support cannot not be disregarded.
References
1 Mathioudakis N et al. An AI-Powered Lifestyle Intervention vs Human Coaching in the Diabetes Prevention Program: A Randomized Clinical Trial. JAMA. 2025. 334(23):2079-89
2 Lalani B et al. Patient engagement, acceptability, and preference of artificial intelligence versus human coaching for diabetes prevention. Digit. Med. 2026. https://doi.org/10.1038/s41746-026-03063-w
Featured image: neirfy on AdobeStock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
