Seeing the Heart through the Eye

CKD is a major public health problem, affecting approximately 850 million people worldwide, roughly twice the number of individuals living with diabetes (422 million).1 Patients with CKD face an elevated risk of cardiovascular disease (CVD) including coronary artery disease, stroke, heart failure, arrhythmias, and sudden cardiac death. The prevalence of CVD rises with CKD severity, from 63.4% in stages G1–G2 and 66.6% in stage G3 to 75.3% in advanced stages G4–G5, with cardiovascular mortality accounting for nearly half of deaths in the most advanced stages.2 Hence, early identification of patients with CKD at risk of developing CVD and timely initiation of evidence-based, disease trajectory-modifying therapy is crucial. Although CKD is a well-established risk factor for CVD, accurately predicting which patients are most at risk remains challenging. Given that cardiovascular morbidity and mortality disproportionately affect individuals with CKD, risk prediction tools derived from the general population may underestimate the risk of atherosclerotic CVD or heart failure in this group.3 The Kidney Disease Improving Global Outcomes 2024 guidelines for CVD risk prediction in patients with CKD recommend two approaches: using existing models with added kidney-specific measures such as eGFR and albuminuria or developing entirely new models exclusively for CKD populations.1 Although “CKD Add-on” models, including eGFR and albuminuria to existing CVD models (pooled cohort equations/Systematic COronary Risk Evaluation), have shown improved prediction, the recently developed Predicting Risk of Cardiovascular disease EVENTs (PREVENT) model, which included eGFR in the primary model rather than as an add-on, has also demonstrated enhanced prediction. However, few models have been developed and validated solely within CKD populations. In this issue, Dhamdhere et al.3 presented a model for predicting CVD risk in patients with CKD using retinal vascular features obtained through deep learning–based segmentation and classification. Microvascular disease plays a key role in the development of both CKD and CVD.1 The retina, with its high vascularity and noninvasive accessibility, serves as a surrogate for systemic vascular health. Retinal vascular parameters such as changes in arteriolar and venular diameters, retinopathy signs, and structural alterations, such as vessel tortuosity or abnormal fractal dimensions, have been shown to be associated with both CKD and CVD. Retinal biomarkers obtained noninvasively through ocular imaging such as fundus photography, optical coherence tomography, or optical coherence tomography angiography (OCTA) collectively referred to as oculomics have been shown to predict systemic diseases.4 Previously, retinal parameters required semiautomated measurement with manual correction, but advances in retinal imaging and artificial intelligence now allow automatic extraction of retinal features which can be used to predict CVD5 or enable direct prediction from retinal images through pattern recognition, without the need for explicit feature extraction, for the assessment of CKD6 or CVD risk factors.7 Cheung et al. showed that deep learning–derived retinal arteriole and venule were predictive of CVD.5 Building on this, Lim et al. demonstrated that these retinal measures were independently associated with incident CVD in individuals with CKD and that adding retinal and kidney function parameters to traditional risk models improved discrimination and reclassification, suggesting that retinal imaging may provide incremental value for CVD risk prediction in those with CKD.8 Extending this work, Dhamdhere et al.3 developed CARE-CKD model (MCARE), a prognostic model for predicting CVD risk in patients with mild to moderate CKD from the Chronic Renal Insufficiency Cohort study in the United States. Using a computer vision pipeline, the team quantified a comprehensive array of retinal vessel features such as density, length, branch angles, and tortuosity from which the top eight were used to train an elastic net model to predict major adverse cardiovascular events (MACE). The performance of MCARE was compared with Framingham Risk Score as well as the PREVENT equation. MCARE showed modestly better discrimination (C-index 0.70 versus 0.66 for Framingham Risk Score and 0.65 for PREVENT) and provided improved risk stratification among patients already classified as “high risk” by these scores. Shapley additive explanations which provide insights into model behavior revealed that denser and longer arterial networks with wider artery and vein branch angles were associated with lower CVD risk, suggesting more effective oxygen delivery. Conversely, greater venular tortuosity predicted higher risk, consistent with endothelial dysfunction and inflammation in CKD. These findings support the concept of the retina as a noninvasive window into systemic microvascular health. This retinal image-based deep learning approach offers several practical advantages for predictive modeling. First, retinal imaging, increasingly available through nonmydriatic fundus cameras in primary care and community settings, provides vascular information often missed by traditional assessments, making it a convenient, noninvasive tool for CVD risk screening. Second, the MCARE model provides an interpretable, feature-based output that may facilitate clinician acceptance and integration into shared decision-making by enabling clinicians to better understand the rationale underlying algorithmic decisions. Third, by restratifying high-risk patients into more granular categories, MCARE could identify those most likely to benefit from aggressive preventive measures such as intensified BP control, lipid lowering, or novel anti-inflammatory agents. The notion of combining retinal imaging with CKD-specific clinical markers in a nomogram, as proposed by the authors, could be embedded into electronic health records to generate individualized CVD risk profiles during routine visits, without the need for invasive testing. Nevertheless, several caveats remain. Although the authors assessed deaths, CVD mortality was not included in the MACE outcome, potentially underestimating cardiovascular risk. The lack of external validation limits both the generalizability of the findings and their broader applicability to other populations or clinical settings. In addition, the model was developed in patients with mild to moderate CKD, and its performance must be evaluated in advanced CKD populations, including those on dialysis who are at the highest risk of CVD. Finally, while interpretability is an advantage, the manual and computational steps involved in feature extraction must be robust, reproducible, and standardized across imaging platforms to ensure consistency. Moving forward, it is important that validation efforts include large, multicenter studies across diverse CKD populations and various ethnic groups. Longitudinal imaging could also reveal whether dynamic changes in retinal vasculature correlate with changes in CVD risk, opening the possibility for retinal imaging to serve as a monitoring tool rather than a one-time assessment. Furthermore, multimodal retinal imaging, such as fundus photography combined with optical coherence tomography angiography, which provides high-resolution imaging of the retinal and choroidal microvasculature may enhance cardiovascular risk prediction by identifying novel imaging biomarkers9 for refined risk stratification. Implementation studies are also needed to assess the effect of integrating the risk prediction tool into clinical workflows and patient education, addressing an often-neglected aspect of CKD care.10 Finally, economic evaluation will also be important, if MCARE enables earlier intervention that prevents MACE, it may prove cost-effective by reducing hospitalization and dialysis-related complications. Retinal imaging, being noninvasive and increasingly accessible, holds promise as a practical tool for early cardiovascular risk assessment in CKD. Its integration into optometry or primary care workflows could help identify high-risk individuals without additional invasive procedures, thereby guiding timely intervention. With further validation and careful implementation, retinal microvascular analysis could become a standard component of cardiovascular risk assessment, enabling earlier detection and treatment to help alter the trajectory of CVD in CKD.