Can AI-driven health management prevent stroke? — 26,000 adults followed for 36 months, randomized by community (Beijing, NCT07779668)

1Trials that measure accuracy, and trials that measure onset

Most medical-AI trials measure how well an AI can tell things apart. That is not what this one sets out to measure. It asks whether putting an AI-supported health-management system into communities actually reduces strokes — the outcome itself. Better image reading does not lower incidence if blood pressure and medication do not hold in daily life.

The design is an attempt to cross the long distance between a prediction and an event, using 36 months of follow-up.

2Randomizing communities rather than individuals

That the unit of assignment is the community rather than the individual says a great deal about the character of the design. Health guidance and encouragement spread easily to those nearby, and mixing intervention and control participants in one community lets the effects bleed into each other.

Splitting by community avoids that contamination, and by implication it means the intervention is conceived as the operation of community health services rather than as a prescription to an individual. Wearable data is explicitly limited to when it is available, so the design does not assume that devices will be distributed.

3Among the largest medical-AI trials this site holds

Across the 750 medical-AI clinical trials this site holds as of 2026-08-28, the median planned enrollment is 200. The 26,000 planned here is 130 times that median. Only 50 trials — 6.7 percent — plan 5,000 or more, while 254 trials, or 34 percent, plan 100 or fewer. In that distribution this one stands out sharply.

Clinical evaluation of medical AI is dominated by small accuracy studies, and a trial that goes after incidence at scale is exceptional. What the AI-ExpoStroke model contains, and whether strokes are in fact reduced, are not part of this registry record.

4Trials that measure accuracy, and trials that measure incidence

Most medical AI trials measure how correctly the AI distinguishes. What this trial measures is the outcome itself.

That randomisation is by community rather than by individual also marks the design. Health guidance and outreach spread easily to those nearby, and mixing intervention and control within one community lets the effects bleed. It means the intervention is conceived not as a prescription to an individual but as the operation of community health itself.