Accuracy of retinal-image artificial intelligence for atherosclerosis on coronary CT and vascular ultrasound: a systematic review and meta-analysis
Abstract
RETROSPECTIVE REGISTRATION. This registration documents the protocol and methods of a completed review; it was created after the review was finished.
Objective
To estimate the accuracy, as the area under the ROC curve (AUC), of artificial intelligence applied to retinal images for detecting atherosclerosis defined by imaging.
Design
Systematic review and meta-analysis of diagnostic test accuracy, reported according to PRISMA 2020, PRISMA-DTA and PRISMA-S. ELIGIBILITY - Population: adults without known atherosclerotic cardiovascular disease (studies excluded only when every participant had established disease). - Index test: models with a learned component (deep or machine learning) applied to colour fundus photographs, ultra-widefield images, optical coherence tomography or OCT angiography. - Reference standard and target conditions (primary definitions): coronary artery calcium on CT (CAC > 0), any coronary plaque on CT angiography, and carotid or femoral plaque on ultrasound, in the same participants. - Accuracy from data independent of model training. Excluded: non-original reports, children only, prognostic targets only, and out-of-scope references such as pulse wave velocity or invasive coronary angiography. SEARCH PubMed, Embase and Web of Science Core Collection, from inception to September 2026, without language limits. SELECTION AND DATA EXTRACTION Two independent assessments with adjudication at title-abstract and full-text stages; duplicate extraction with a standardised form; protocol rules to choose one estimate per study and analysis; studies sharing participants contributed one estimate per analysis. RISK OF BIAS QUADAS-3 (version 1.2) with review-specific guidance for AI index tests; two independent assessments and adjudication. SYNTHESIS AUCs on the logit scale pooled with Bayesian random-effects models (normal prior for the mean, half-normal(0.5) prior for between-study heterogeneity), with a frequentist REML model with the Hartung-Knapp-Sidik-Jonkman adjustment as counterpart. Reported: 95% credible and prediction intervals, and posterior probabilities of an AUC above 0.70 and 0.80. Sensitivity analyses: prior scale, leave-one-out, external validation only, alternative estimates, and a multiverse analysis. Exploratory multilevel models compared image-only with image-plus-clinical models. CERTAINTY OF EVIDENCE GRADE for diagnostic test accuracy. AVAILABILITY Protocol, extracted data and analysis code are available from the corresponding author on reasonable request.
