Clinical utility of an AI-based retinal imaging model for cardiovascular risk prediction in hypertensive retinopathy

Purpose

This study presents an independent clinical evaluation of Dr.Noon CVD, a commercially developed artificial intelligence (AI)-based retinal imaging model that estimates cardiovascular disease (CVD) risk. We assessed whether the model can effectively evaluate CVD risk in patients with hypertensive retinopathy (HR), a population in which the applicability of conventional CVD risk models remains uncertain.

Methods

We retrospectively analyzed 102 age-matched patients from Hanyang University Guri Hospital and classified them into normal (Group 1), low-grade HR (Group 2), and high-grade HR (Group 3) groups using Keith–Wagener–Barker grading. CVD risks were assessed via Dr.Noon CVD score and conventional risk scores (PREVENT, PCE, SCORE2, SCORE2-Diabetes). Associations and predictive performance were evaluated using logistic regression, area under the ROC curve (AUC), and net reclassification improvement (NRI).

Results

Low-density lipoprotein cholesterol, estimated glomerular filtration rate, and blood pressure were significantly higher in the HR groups versus Group 1 (p < 0.05). Dr.Noon CVD scores differed significantly across HR grades (p = 0.002), particularly between Groups 3 and 1 (p = 0.002), while conventional scores showed no significant separation (p > 0.1). Higher Dr.Noon CVD scores were associated with Group 3 in both unadjusted (OR = 1.11; p = 0.001) and adjusted models (OR = 1.33; p < 0.001). Scores remained stable between acute and chronic HR (p = 0.966). Combining Dr.Noon CVD score with conventional models improved CVD risk classification (higher AUC and NRI).

Conclusions

Dr.Noon CVD identified elevated CVD risk in HR patients and enhanced stratification when combined with existing models. These results support the use of retinal biomarkers in CVD risk assessment and the need for multidisciplinary management.