An app matched human coaches in a diabetes trial. Both worked about a third of the time.

The Diabetes Prevention Program is one of the few lifestyle interventions in medicine with a serious evidence base behind it, and one of the few that health systems will actually pay for. It is also expensive to deliver, because it is delivered by people.

A randomised trial appearing in this week's issue of JAMA [s4] asked whether an app could do the same job. Its answer, carefully stated, is that it was not shown to be worse.

The design

368 adults with overweight or obesity and prediabetes were randomised at two US clinical centres — in Baltimore, Maryland, and Reading, Pennsylvania — between October 2021 and December 2024 [s2]. Median age was 58; 71% of participants were female [s2].

One group was referred to an AI-powered Diabetes Prevention Program delivered through a mobile app paired with a Bluetooth-enabled scale. The other was referred to a remotely delivered, human-coach-led programme [s2].

The primary outcome was a composite, and its construction matters. A participant met it at 12 months by achieving any one of: at least 5% weight loss; at least 4% weight loss plus 150 minutes of weekly physical activity; or a reduction in HbA1c of at least 0.2 percentage points while remaining below 6.5% [s2]. These are the risk-reduction benchmarks the programme is judged against, not arbitrary thresholds.

The result

At 12 months, 31.7% of those referred to the AI-led programme met the composite, against 31.9% of those referred to human coaching — a difference of 0.2 percentage points [s2].

The trial was designed as a noninferiority study with a margin of 15 percentage points. The lower boundary of the 95% confidence interval on the difference came in at −8.2 percentage points [s2], inside the margin. The authors' conclusion is that referral to the AI-led programme was noninferior to referral to a human-led one [s3].

One secondary finding is more interesting than the primary. Participants referred to the AI programme actually started it more often: 93.4% initiated, against 82.7% in the human-coached arm [s2].

What noninferiority means and does not mean

Three things are worth separating.

First, this is not a finding that the app is as good as a coach. It is a finding that the trial did not detect a difference larger than 15 percentage points, on a composite outcome, in this population. A 15-point margin is wide. It was pre-specified, which is the correct procedure, but a reader should hold it in view: the design would have declared noninferiority even if the app had performed meaningfully worse.

Second, the confidence interval is doing work the point estimate hides. A true difference anywhere from about 8 points worse to some amount better is compatible with the data [s2]. With 368 participants, that width is unavoidable.

Third — and this is the finding that should travel furthest — both arms produced the composite outcome in roughly one participant in three [s2]. Whatever the delivery mechanism, about two-thirds of people referred did not hit the benchmark within a year. That is consistent with what lifestyle intervention looks like at scale, and it is a more important number for anyone planning a programme than the comparison between arms.

The limits

The trial ran at two centres, both in the mid-Atlantic United States [s2], which constrains generalisability. The comparison is between referral to each programme, not between the programmes themselves as delivered — which is the pragmatic question a health system faces, but means the arms differ in uptake as well as content. The initiation gap [s2] suggests the app's advantage may lie substantially in lowering the friction of starting rather than in coaching quality.

Twelve months is also short for an outcome that matters. Weight regain and HbA1c drift are well-documented beyond the first year, and this trial does not speak to either.

The paper was published online in October and appears in the 16 December issue [s1][s4].

Why it matters anyway

The economics are the reason this trial exists. Human-delivered prevention programmes are constrained by coach supply; an app is not. If an automated programme can reach parity — or something inside a defensible margin of parity — the ceiling on how many people can be enrolled changes.

That is a real result. It is a different claim from "AI is as good as a human", and the trial does not support the stronger version.

What to watch

Whether longer follow-up is published, whether the finding replicates in populations less like this one, and whether payers treat noninferiority on a 15-point margin as sufficient for coverage. The third question will be settled well before the first two.

This article is informational and is not medical advice.