Fundus Photograph-Derived Computational Features Predict Risk of Cardiovascular Events in the Chronic Renal Insufficiency Cohort Clinical Observational Study

Key Points Retinal imaging reveals microvascular changes associated with cardiovascular (CV) risk in patients with CKD. A retinal vascular model improves risk stratification compared with established clinical risk scores for CV events. Integrating retinal analysis with clinical markers enhances CV risk assessment in patients with CKD. Background Patients with CKD face an elevated but variable risk of cardiovascular (CV) disease. Retinal imaging in CKD provides a noninvasive opportunity for CV risk stratification through microvascular analysis. The objective of this study was to evaluate retinal vascular features extracted via computer vision and machine learning approaches for CV risk and their added value over established risk calculators in patients with CKD. Methods Retinal scans from 1333 participants of the multicenter clinical observational study Chronic Renal Insufficiency Cohort (NCT00304148) were analyzed. A deep learning pipeline segmented vessels and then identified arterioles and venules from them. Segmented vessel, arteriole, and venule masks were used to extract 384 vascular features. An elastic-net model—Cardiovascular Assessment through Retinal Evaluation in CKD (CARE-CKD; M CARE )—was trained, using the top eight features, on 567 participants (101 major adverse cardiovascular events [MACE]: composite of myocardial infarction, stroke, heart failure) and validated on 244 participants (44 MACE). A nomogram integrating M CARE with clinical markers (age, sex, blood pressure, smoking, eGFR, albuminuria, cholesterol, body mass index, and diabetes status) was developed. Results M CARE demonstrated strong prognostic performance for predicting MACE, (C-index=0.70, hazard ratio [HR] =3.95, above versus below median; 95% confidence interval [CI], 2.36 to 6.63; P <0.001), outperforming the Framingham Risk Score (C-index=0.66; HR=1.06; Likelihood Ratio Test [LRT] P < 0.01) and Predicting Risk of cardiovascular disease EVENTs (PREVENT; C-index=0.65; HR=1.84; LRT P < 0.001) calculators. M CARE improved risk stratification within Framingham Risk Score–based high-risk (HR=3.73; P < 0.001) and Predicting Risk of cardiovascular disease EVENT–based high-risk (HR=4.73; P < 0.001) categories. Nomogram enhanced risk stratification (C-index=0.77; HR=3.81; P <0.0001) compared with clinical markers (LRT P < 0.01). Conclusions CARE-CKD provides a novel, opportunistic approach to CV risk assessment in CKD, outperforming the established risk calculators and refining stratification within high-risk categories. By enabling earlier identification, close monitoring, and targeted management of high-risk patients, CARE-CKD addresses gaps left by traditional calculators, maximizing the benefits of emerging therapies and potentially improving long-term outcomes.