A retinal imaging–based model showed moderate-to-good discrimination for detecting coronary artery disease and performed better when combined with clinical risk factors in a retrospective study of patients undergoing coronary angiography.

Researchers analyzed data from 417 patients with suspected angina at a single center in China who underwent first-time coronary angiography. Fundus photography was performed within 24 hours prior to angiography, and coronary artery disease (CAD) was defined as at least 50% stenosis in one or more coronary vessels.

The combined model, incorporating both retinal imaging features and clinical variables, achieved an area under the receiver operating characteristic curve (AUROC) of 0.802, with a sensitivity of 0.797 and specificity of 0.679. By comparison, the AUROC was 0.748 for a model based on clinical risk factors alone and 0.694 for a model based on retinal data alone.

Key Findings

Quantitative retinal vascular parameters were independently associated with CAD after multivariable adjustment, including fractal dimension and vessel density, along with selected measures of vessel diameter and morphology.

In the final combined model, decreased fractal dimension, reduced optic disc axis ratio, and shorter optic disc-to-macula distance were the retinal variables retained as independent predictors. These were identified alongside clinical factors including male sex, dyslipidemia, and higher levels of lipoprotein(a), triglycerides, low-density lipoprotein cholesterol, and glycated hemoglobin.

Clinical Context

The retinal microvasculature offers a noninvasive window into systemic vascular health, sharing structural and pathophysiological characteristics with coronary vessels. These similarities have led to growing interest in retinal imaging as a potential tool for cardiovascular risk assessment.

Unlike some deep learning approaches, the model used explicitly quantified retinal features, improving interpretability. The use of non-mydriatic fundus photography may also support feasibility in routine clinical workflows.

The model demonstrated higher sensitivity than specificity, suggesting it may be more useful for identifying patients at increased risk of CAD than for ruling out disease.

Limitations

The findings are limited by the study’s retrospective, cross-sectional, single-center design, which precludes causal inference and may limit generalizability. The cohort consisted of patients referred for coronary angiography, representing a high-risk population rather than a general screening group.

In addition, the study population was predominantly Han Chinese, and the model did not incorporate a broader range of biomarkers such as inflammatory or hematologic indices, further limiting generalizability.

Bottom Line

Retinal vascular phenotyping may provide a useful adjunct to traditional clinical risk assessment for CAD, particularly when combined with established risk factors. However, further validation in larger, more diverse populations is needed before broader clinical application.

One co-author (Saiguang Ling) is affiliated with Evision Technology, the company that developed the EVisionAI fundus analysis system used in this study. The authors declared no competing interests.

Source: BMJ Open