AI-powered cardiometabolic assessment from a single facial video

Introduction

Hypertension and diabetes are leading modifiable risk factors for cardiovascular disease, yet a substantial proportion of affected individuals remain undiagnosed. Current screening relies on dedicated clinical encounters or wearable devices, both of which limit population-level coverage. A passive, contactless monitoring system that can be embedded in everyday environments could overcome these barriers and enable early detection at scale.

Purpose

We aimed to develop and evaluate a machine-learning algorithm that detects hypertension and diabetes and estimates blood pressure from short-duration facial video captured by a spectroscopic camera, without any physical contact or blood sampling.

Methods & Results

In this single-center prospective cohort study, 285 consenting participants (patients and healthy volunteers) underwent high-speed spectroscopic video recording of the face and hands. Concurrent cuff-based blood pressure measurements served as the reference standard for hypertension (systolic ≥130 mmHg or diastolic ≥80 mmHg, per AHA 2017 criteria), while HbA1c values and physician-diagnosed diabetes (≥6.5% or clinical diagnosis) were used as the reference for diabetes.

The algorithm extracted three categories of features from the video data: pulse wave dynamics, skin blood-flow patterns, and spectral characteristics. For hypertension detection, the pulse wave-based analysis achieved 95.0% accuracy using a 30-second recording in an ensemble model (normal blood pressure sensitivity 100.0 ± 0.00%, hypertension sensitivity 89.2 ± 2.30%), and 90.3% accuracy using 5 seconds of data (normal blood pressure sensitivity 98.3 ± 6.22%, hypertension sensitivity 84.5 ± 7.2). For diabetes detection, a model using 30-second videos based on blood flow patterns achieved 88.2% accuracy (Fine Gaussian SVM, ROC-AUC 0.97), and a model using 5-second videos achieved 81.2% accuracy (SVM kernel, ROC-AUC 0.98). Five-fold cross-validation was used for these binary classifications. In addition, for blood pressure estimation using 30-second videos of the face only, a model using three parallel models—Random Forest, XGBoost, and LightGBM—achieved a Mean Absolute Percentage Error (MAPE) of 9.7% and a model correlation coefficient of 0.968.

Conclusion

A machine-learning algorithm applied to spectroscopic facial video can detect hypertension and diabetes with high accuracy from recordings as short as 5 seconds, and can estimate blood pressure without a cuff. These findings suggest that contactless biometric sensing may serve as a scalable screening tool for cardiovascular risk factors, warranting further validation in larger, multi-center cohorts.ESC2026