Breaking News:AI Retinal Imaging for Blood Pressure Biometry– What Just Happened

A NEW study suggests that blood pressure biometry may soon be possible without cuffs or physical contact, using AI to analyse images of the retina. Researchers report that a deep learning model can estimate systolic and diastolic blood pressure from non-mydriatic fundus photographs with accuracy comparable to standard arm cuff measurements.

Why Blood Pressure Biometry Needs Improvement

Blood pressure remains one of the most widely used clinical indicators, yet traditional cuff measurements capture only a brief snapshot that can vary significantly between readings. Systolic values can fluctuate by as much as 10 mm Hg on consecutive measurements due to stress, posture, time of day, and the so-called white coat effect. These limitations mean that short-term readings may poorly reflect a patient’s long-term hypertensive burden. Because chronic hypertension leaves lasting structural changes in the retinal microvasculature, researchers have increasingly explored whether retinal imaging could offer a more stable and biologically meaningful measure of blood pressure status.

Study Design and AI Model Performance

The study used more than 105,000 retinal images from over 51,000 participants in the UK Biobank. Researchers trained a convolutional neural network to predict systolic and diastolic blood pressure using paired fundus images and arm cuff readings taken during the same visit. Cuff measurements served as the reference standard. The model achieved a mean absolute error of 9.81 mm Hg for systolic pressure and 6.00 mm Hg for diastolic pressure, with coefficients of determination of 0.36 and 0.30, respectively. These error ranges were comparable to the known variability of manual cuff measurements, and substantially better than predicting blood pressure using population averages alone.

Implications for Chronic Blood Pressure Assessment

Model visualisation suggested the AI relied primarily on retinal blood vessel architecture and optic disc features, supporting the idea that retinal structure reflects long-term vascular stress rather than transient fluctuations. The researchers argue that retinal blood pressure biometry may function similarly to HbA1c testing in diabetes, offering insight into chronic physiological status rather than momentary readings. While additional clinical trials are needed to validate performance in real-world settings, the findings indicate that non-contact retinal imaging could eventually complement or enhance existing blood pressure monitoring, particularly for identifying sustained hypertension and microvascular risk.

Reference

Bressler I et al. Non-contact optical blood pressure biometry using AI-based analysis of non-mydriatic fundus imaging. BMJ Innovations. 2026;bmjinnov-2025.