Retinal image based artificial intelligence biomarker for cardiovascular risk and cross-sectional association with coronary atherosclerosis
S L Papadopoulou, K Michailidis, S Iordanidis, V Liakopoulou, I Styliadis, D Nam, J Seo, S Nusinovici, S Thakur, T H Rim, V SachpekidisAbstract
Background
Artificial intelligence (AI) based retinal image analysis has been developed as a non-invasive biomarker for prediction of future cardiovascular risk. Its cross-sectional performance against current coronary imaging markers of atherosclerosis requires evaluation in independent populations.
Purpose
To assess the cross-sectional association between an AI-derived retinal cardiovascular risk biomarker and coronary atherosclerosis defined by Coronary Artery Calcium (CAC), Coronary Artery Disease Reporting and Data System (CAD-RADS) classification, and plaque burden on Coronary Computed Tomography Angiography (CCTA) in a Greek population.
Methods
In a cohort of 1,233 individuals undergoing CCTA, same-day retinal photographs were analysed to generate categorical cardiovascular risk groups (low, moderate, high) using the AI-based retinal biomarker. Although designed for future risk prediction, performance was evaluated cross-sectionally by comparing low retinal risk versus combined moderate or high retinal risk for detection of coronary atherosclerosis on CCTA. Sensitivity (SN), specificity (SP), and discrimination using area under the receiver operating characteristic curve (AUC) were calculated for CAC >0, CAD-RADS >0 and plaque burden ≥2.
Results
Median age was 54 years and 66.1% were male. The proportion classified as high risk by the AI-derived retinal biomarker increased across CAC categories, from 28.5% in CAC 0 to 54.2% in CAC 1–99 and 71.1% in CAC ≥100. A similar gradient was observed across CAD-RADS categories, with high retinal risk in 27.6% of CAD-RADS 0, 49.7% of CAD-RADS 1, 68.6% of CAD-RADS 2, 64.5% of CAD-RADS 3, and 69.2% of CAD-RADS 4–5. Likewise, as plaque burden severity increased, a greater proportion of individuals were classified as high retinal risk. Performance for detection of CAC >0 was SN 94%, SP 25.7%, AUC 0.737. For CAD-RADS >0, SN 92%, SP 26.9%, AUC 0.719. For plaque burden ≥2, SN 96.3%, SP 21.2%, AUC 0.727.
Conclusions
An AI-based retinal biomarker developed for future cardiovascular risk prediction demonstrates high sensitivity with low specificity for cross-sectional detection of imaging-defined coronary atherosclerosis in a Greek population. Consistent retinal risk gradients across CAC, CAD-RADS, and plaque burden indicate concordance with coronary disease severity and support its potential role as a non-invasive risk-stratification tool to guide further cardiovascular evaluation. Prospective studies are needed to establish its clinical utility within preventive care pathways.Figure 1