Camera-based screening approach using zygomatic and cheek regions performs comparably to existing non-invasive diagnostic methods
Deep learning analysis of facial images identified hypertension with 83% accuracy, with specific facial regions performing nearly as well, according to a study published in Scientific Reports.
The study, led by Jing Wang, PhD, of Beijing University of Chinese Medicine, analyzed facial images from 375 patients with hypertension and 131 normotensive controls. Notably, models using only the zygomatic or cheek regions each achieved 82% accuracy—performing on par with the full-face model.
The approach performed comparably to existing non-invasive methods. Traditional statistical models based on facial color achieved area under the curve (AUC) values around 82% to 83%, while contact-based photoplethysmography methods reported accuracy of approximately 73%. The proposed framework achieved 83% accuracy, with an F1-score of 0.75 and AUC of 84%.
“This study demonstrates that deep learning analysis based on facial images can serve as a scalable, passive, non-invasive initial screening tool, operable in everyday environments using only standard cameras,” the researchers wrote. “Notably, the zygomatic and buccal regions exhibit specificity in identifying hypertension.”
According to the researchers, the method could address key barriers in hypertension detection, including low screening adherence, asymptomatic disease onset, and measurement biases such as white-coat hypertension. Unlike blood pressure cuffs or photoplethysmography sensors, the approach requires only standard cameras and works in everyday environments.
The researchers employed an improved U-Net model to segment faces into six anatomically defined regions, then trained ResNet-based classifiers to predict hypertension. The segmentation model achieved a mean Intersection over Union of 98%.
Among individual facial regions, the forehead achieved 78% accuracy, the nose reached 77%, the jaw achieved 77%, and the lip region achieved 73%. Grad-CAM visualization showed the models primarily focused on the zygomatic and cheek regions when distinguishing patients with hypertension from controls—areas that correspond to facial manifestations emphasized in Traditional Chinese Medicine.
Participants were recruited from Dongzhimen Hospital between August and December 2023. Patients with hypertension showed systolic blood pressure of 140 mmHg or greater and/or diastolic blood pressure of 90 mmHg or greater on multiple measurements. Facial images were captured under standardized lighting using specialized diagnostic imaging systems.
However, the sample size remains relatively small for deep learning applications, potentially limiting generalizability. The researchers noted that future work should incorporate larger, multicenter, and more diverse cohorts to ensure applicability across different ethnicities, genders, and age groups.
The researchers noted that the approach is intended as a complementary screening tool rather than a replacement for established clinical blood pressure measurements.
Disclosures can be found in the study.
Source: Scientific Reports