AI-based diagnosis of hypertension and diabetes from a single facial video

A study to be presented at ESC Congress 2026 reveals that AI analysis of facial videos can accurately diagnose hypertension and diabetes from a single recording. Approximately one in seven adults worldwide have hypertension, and a sixth have diabetes. Current screening methods are limited by the need for clinical visits or wearable devices.

Researchers at the University of Tokyo sought to develop an AI algorithm that could perform contactless screening in everyday settings to detect these conditions earlier and on a larger scale.

The algorithm analyzed spectroscopic facial videos and palm recordings, extracting data on pulse-wave dynamics, skin blood flow patterns, and skin coloring characteristics. In a prospective single-center study, 215 participants, including diagnosed patients and healthy volunteers, underwent short video recordings. The algorithm achieved 95.0% accuracy in detecting hypertension from a 30-second recording, with 89.2% sensitivity and 90.3% accuracy from a 5-second recording.

For diabetes detection, the algorithm reached 88.2% accuracy from a 30-second video and 81.2% from a 5-second video.

Notably, the algorithm could estimate blood pressure from facial videos alone, with a mean absolute percentage error of 8.6% for systolic blood pressure (SBP). While the standard deviation error of ±12.0 mmHg exceeded the AAMI's limit of ±8.0 mmHg, the algorithm still met the criterion of ±5.0 mmHg. The researchers plan to validate these findings in larger, multicenter studies to support real-world application.

If validated, this contactless approach could enable people to be screened in everyday settings without cuffs, blood sampling, or clinic visits, potentially identifying at-risk individuals who would otherwise remain undiagnosed and untreated.