Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs.
Level 3 - non-randomized controlled study
Retrospective and prospective observational cohort validation studies across multiple independent cohorts
PubMed 33890578 · doi:10.1016/S2589-7500(21)00043-1
What was done
Researchers developed a deep-learning algorithm (RetiCAC) using 216,152 retinal photographs across five datasets from South Korea, Singapore, and the UK. The model was trained on data from a South Korean health-screening centre to predict the probability of coronary artery calcium (CAC) presence. RetiCAC scores were divided into tertiles and validated for cardiovascular event prediction using Cox proportional hazards models on external test cohorts: a South Korean clinical cohort (n=527, 5-year follow-up), a Singapore population-based cohort (n=8,551, 10-year follow-up), and the UK Biobank (n=47,679, 10-year follow-up). Incremental value when added to the Pooled Cohort Equation (PCE) was evaluated in the UK Biobank.
What was found
RetiCAC predicted CAC presence with an area under the receiver operating characteristic curve (AUROC) of 0.742 (95% CI 0.732–0.753). In the South Korean cohort (33/527 events), the three-strata RetiCAC matched CT-measured CAC risk stratification with a concordance index of 0.71. In the Singapore cohort (310/8,551 fatal events), RetiCAC was associated with fatal cardiovascular events (hazard ratio [HR] trend 1.33, 95% CI 1.04–1.71). In the UK Biobank (337/47,679 fatal events), adding RetiCAC to the PCE improved risk stratification in intermediate-risk (HR trend 1.28, 95% CI 1.07–1.54) and borderline-risk groups (HR trend 1.62, 95% CI 1.04–2.54), with a continuous net reclassification index of 0.261 (95% CI 0.124–0.364).
Why it matters
Retinal photography combined with deep learning offers a non-invasive surrogate for CT-derived coronary artery calcium scoring that incrementally improves standard cardiovascular risk stratification equations, which could facilitate screening in low-resource settings.
Limits
The abstract lacks demographic details beyond study country origins. Event rates in some validation cohorts were low (0.7% in the UK Biobank cohort), the AUROC for CAC detection was modest (0.742), and the abstract provides no prospective clinical utility or implementation data.
Cited by
- supports AI analysis of retinal imaging can assess risk of coronary artery disease, stroke, and predict coronary artery calcium score.