AI detects hypertension and diabetes from short videos, in study

A short, high-speed recording of a person’s face and palms produced two medical signals in a Japanese study: 95.0% accuracy for detecting hypertension from 30 seconds of video, and 88.2% accuracy for diabetes from facial blood-flow patterns over the same duration. The work from the University of Tokyo and Institute of Science Tokyo will be presented at ESC Congress 2026.

The prospective, single-center study included 215 participants, among them diagnosed patients and healthy volunteers. The machine-learning system analyzed pulse-wave dynamics — signals linked to artery stiffness — along with skin blood-flow patterns and the spectral characteristics of skin coloring. Participants also underwent conventional assessments for hypertension and diabetes.

Speed was central to the test. Hypertension accuracy remained 90.3% with a 5-second video, while diabetes detection reached 81.2%. For hypertension, the strongest reported result used pulse-wave analysis from both facial and palm video. The system could also estimate systolic blood pressure from facial video alone, without a cuff.

That blood-pressure estimate shows both the promise and the boundary. The mean error was −2.6 mmHg, within the Association for the Advancement of Medical Instrumentation’s limit of ±5.0 mmHg, but the standard deviation error was ±12.0 mmHg, above its ±8.0 mmHg criterion. The researchers plan larger, multicenter datasets and further feature optimization to reduce that spread.

So what changes in practice? If the findings hold in larger and more diverse cohorts, screening could move into everyday settings where a cuff, blood sample or dedicated clinic visit is impractical. That could help identify people who currently remain undiagnosed and untreated. For now, the evidence is from one center, and the team is still describing a route to real-world application rather than a deployed medical service.