Pivotal trial of a deep-learning-based retinal biomarker (Reti-CVD) in the prediction of cardiovascular disease: data from CMERC-HI.
Level 3 - non-randomized controlled study
Retrospective cohort validation study of a prognostic biomarker algorithm with 5-year follow-up
PubMed 37847669 · doi:10.1093/jamia/ocad199
What was done
Researchers evaluated Reti-CVD, a deep-learning artificial intelligence software as a medical device (AI-SaMD) that predicts cardiovascular disease (CVD) risk from retinal images, using retrospective data from 1,106 participants in the Cardiovascular and Metabolic Diseases Etiology Research Center-High Risk (CMERC-HI) cohort. Participants were categorized into low-, moderate-, and high-risk tiers. Cox proportional hazard models evaluated the association between Reti-CVD risk tiers and 5-year CVD events, adjusting for traditional risk factors and vascular biomarkers including coronary artery calcium (CAC), carotid intima-media thickness (CIMT), and brachial-ankle pulse wave velocity (baPWV).
What was found
Over 5 years of follow-up, 33 participants (3.0%) experienced CVD events. Across the three Reti-CVD risk tiers, the hazard ratio trend for CVD events was 2.02 (95% CI, 1.26-3.24). In a multivariable model incorporating Reti-CVD, CAC, CIMT, baPWV, and traditional risk factors, Reti-CVD risk tiers were the only variables significantly associated with increased CVD risk relative to the low-risk reference group (moderate risk: HR 2.40 [95% CI, 0.82-7.03]; high risk: HR 3.56 [95% CI, 1.34-9.51]).
Why it matters
The study provides clinical validation data supporting regulatory authorization of an automated, non-invasive retinal biomarker for cardiovascular risk stratification alongside standard vascular assessments.
Limits
The analysis is retrospective and limited to a single high-risk cohort (CMERC-HI). The absolute number of cardiovascular events was small (33 events among 1,106 participants), leading to wide confidence intervals and imprecision in multivariable estimates (such as the moderate-risk estimate crossing unity). Generalizability across unselected populations and diverse ethnicities was not assessed.
Cited by
- supports AI analysis of retinal imaging can assess risk of coronary artery disease, stroke, and predict coronary artery calcium score.