**Background:** Cardiovascular disease (CVD) remains the leading cause of death globally. Noninvasive assessment of microvascular health is challenging, but the retinal vasculature—accessible via fundus photography, OCT angiography (OCT-A), and adaptive optics—offers a unique window into systemic vascular status. Over the past decade, 'oculomics' has emerged, applying artificial intelligence (AI) to retinal images to extract microvascular biomarkers for cardiovascular risk prediction. This narrative review summarizes developments up to August 2022.
**Methods:** A PubMed/Medline search was conducted combining ophthalmology terms (e.g., 'retina', 'fundus photographs', 'OCT-A', 'oculomics') with AI terms ('artificial intelligence', 'deep learning', 'convolutional neural network'). English peer-reviewed articles deemed relevant by title/abstract inspection were included. The review is not systematic and did not follow a full PRISMA checklist.
**Key Results:** The review covers three main areas of AI application:
1. **Prediction of cardiovascular risk factors:** Poplin et al. used a CNN on UK Biobank and EyePACS datasets to predict age (MAE 3.26 years, 95% CI 3.22–3.31), sex (AUC 0.97, 95% CI 0.96–0.97), smoking status (AUC 0.71, 95% CI 0.70–0.73), and systolic blood pressure (MAE 11.35 mmHg, 95% CI 11.18–11.51). Kim et al. reported age prediction with MAE 3.06 years (95% CI 3.03–3.09) using ResNet-152 on 24,366 fundus images. Cheung et al. developed SIVA-DLS, an automated retinal vessel calibre measurement tool with high agreement to semi-automated software (ICCs 0.82–0.95).
2. **Prediction of CVD biomarkers:** Son et al. developed a DL algorithm to discriminate high (≥100) vs. zero coronary artery calcium (CAC) scores, achieving AUC 0.823 (unilateral) and 0.832 (bilateral fundus images); combining with clinical risk factors improved AUC to 0.886. Rim et al. (RetiCAC) showed that a DL-based CAC score from fundus photos could further stratify intermediate-risk groups by Pooled Cohort Equations (PCE), with HR 1.62 (95% CI 1.04–2.54) for SCORE cardiovascular events in the borderline-risk group. Chang et al. developed DL-funduscopic atherosclerosis score (DL-FAS) to predict abnormal carotid intima-media thickness; in low and intermediate Framingham Risk Score subgroups, the highest DL-FAS tercile had significantly higher cardiovascular mortality risk (HR 4.76, 95% CI 1.05–21.63; and HR 3.14, 95% CI 1.04–9.47, respectively).
3. **Prediction of major cardiovascular events:** Cheung et al. found narrower central retinal arteriolar equivalent (HR per SD decrease 1.13, 95% CI 1.02–1.26) independently associated with incident CVD events after multivariable adjustment. Poplin et al. predicted 5-year MACE with AUC 0.70 (95% CI 0.65–0.74) from fundus images alone, comparable to SCORE risk (AUC 0.72, 95% CI 0.67–0.76); combining both did not significantly improve prediction. Kawasaki et al. used a two-stage multitask DL network achieving AUC 0.738 (95% CI 0.710–0.766) from retinal images alone for 5-year MACE, outperforming SCORE/SCORE2 and FRS. Nusinovici et al. (RetiAGE) reported that retinal-age prediction independently predicted cardiovascular mortality (HR 2.42, 95% CI 1.69–3.48, comparing Q4 vs. Q1) after adjusting for chronological and phenotypic age, though incremental improvement in model performance was minimal (0.08–1.8%).
**Clinical Implications:** AI-based retinal biomarkers offer a noninvasive, potentially low-cost tool for cardiovascular risk stratification. They may add incremental value to traditional risk scores (FRS, PCE, SCORE) in specific subgroups, particularly intermediate-risk patients. However, the review identifies several limitations: (1) many studies use extreme cut-off values (e.g., CAC >100 vs. 0) not reflective of clinical workflows; (2) the incremental value over existing models remains unproven—Poplin et al. found no improvement when combining retinal algorithm with SCORE; (3) methodological issues include lack of external validation, non-standardized reporting, and the 'black box' problem of DL models; (4) cost-effectiveness and real-world integration into healthcare workflows have not been studied. The authors conclude that oculomics shows promise but requires interventional studies demonstrating that AI-based retinal biomarkers can motivate therapy changes (e.g., statin initiation) before clinical adoption.