Retinal vascular phenotyping for early detection of coronary artery disease: quantitative assessment and diagnostic modelling

OBJECTIVES: To investigate the association between quantitative retinal vascular parameters and coronary artery disease (CAD) and to evaluate the efficacy of a retinal phenotype-based diagnostic model as a non-invasive tool for early CAD screening. DESIGN: A retrospective cross-sectional study. SETTING: A single-centre study conducted at the Cardiovascular Center of Beijing Tongren Hospital, Capital Medical University, China, between January and October 2024. PARTICIPANTS: 417 patients with suspected angina undergoing their first coronary angiography (CAG) were enrolled. Inclusion criteria were age >18 years and high-quality fundus photography within 24 hours pre-CAG. Major exclusions were prior coronary interventions, severe systemic/valvular heart diseases and ocular conditions impairing retinal vascular visualisation. PRIMARY AND SECONDARY OUTCOME MEASURES: The primary outcome was the association between quantitative retinal vascular parameters and the presence of CAD (defined as ≥50% stenosis). Secondary outcomes included the diagnostic performance area under the receiver operating characteristic curve (AUROC) of three predictive models: one based on quantitative retinal vascular parameters alone, one based on traditional risk factors and a combined model integrating both retinal and clinical variables. RESULTS: This study enrolled 417 patients undergoing initial CAG. Compared with non-CAD controls (n=190), patients with CAD (n=227) had higher prevalence of hypertension, dyslipidaemia and diabetes, along with elevated levels of fasting blood glucose, lipoprotein(a) (Lp(a)), triglyceride (TG) and glycated haemoglobin (HbA1c) (all p<0.05). Quantitative fundus analysis revealed that multiple retinal vascular parameters were independently associated with CAD after multivariable adjustment, including fractal dimension (FD), vessel density (VD) and specific zonal measures of vessel diameter and tortuosity (all p<0.05). Multivariable logistic regression incorporating both fundus and clinical variables identified the following independent predictors of CAD: a decrease in FD (OR=0.26, 95% CI 0.16 to 0.41, p<0.01), reduced optic disc long-to-short axis ratio (OR=0.04, 95% CI 0.004 to 0.46, p=0.01) and optic disc-to-macula distance (OR=0.91, 95% CI 0.86 to 0.97, p<0.01), male sex, dyslipidaemia and elevated levels of Lp(a), TG, low-density lipoprotein cholesterol and HbA1c (all p<0.05). The final diagnostic model achieved an AUROC of 0.802 (95% CI 0.76 to 0.845), with a sensitivity of 0.797 and a specificity of 0.679 at the optimal cut-off. Internal validation via bootstrap resampling (1000 iterations) confirmed the robustness of the identified predictors. CONCLUSION: Our findings, derived from an artificial intelligence-based fully automated quantitative retinal vascular parameters measurement method, revealed that multiple quantitative fundus parameters-including FD, VD and other morphological parameters were significantly associated with CAD risk. The CAD diagnostic model we developed demonstrates strong performance and high interpretability, making it suitable for early CAD screening and diagnosis.