Key points are not available for this paper at this time.
Coronary artery disease (CAD) remains the leading cause of death and current screening methods are limited. Color fundus photography (CFP) has been explored in literature mostly on the basis of associations and exploratory deep learning approaches with only indirect end points. In this explorative study, we aimed to assess the predictive abilities and limitations of explainable CFP-based features using machine learning and same-visit coronary angiography (CA) outcomes as end points for the first time. Patients undergoing CA were imaged with CFP (Zeiss Clarus 500, Zeiss, Oberkochen, Germany) during the same visit. Coronary plaque burden was assessed using the Gensini Score. Retinal features were extracted using Automorph. Machine learning models were trained and evaluated using fivefold cross-validation. Shapley additive explanations (SHAP) values quantified feature importance and interactions. Of 977 screened patients, 632 (1293 eyes) were assessed. CFP features alone reached moderate predictive performance (area under the receiver operating characteristic (AUROC) 0.692). Adding dimensionally reduced CFP features to clinical baselines consistently improved performance, with the best configuration yielding an AUROC 0.775, average precision (AP) 0.752, and Brier score 0.203. Net reclassification improvement (NRI)/integrated discrimination improvement (IDI) analyses supported improved reclassification for age + sex and basic clinical baseline models, but not for extended clinical baseline models. SHAP analysis revealed vessel width, density, and tortuosity as important vascular retinal indicators of CAD burden. Interaction analysis revealed nonlinear, age-, sex-, and diabetes-dependent effects. CFP features modestly improved the CAD classification beyond clinical baselines. Our findings illustrate the potential and the limitations of CFP features and indicate the need for complex modeling, methodological improvement, and multimodal approaches to achieve valuable classification efficacy.
Jeremic et al. (Sat,) studied this question.