Abstract
BACKGROUND: Coronary artery disease (CAD) is a leading cause of morbidity and mortality globally. However, current detection methods have various safety concerns and are not suitable for all populations. Exploring safe, noninvasive detection methods is crucial. OBJECTIVES: The objective of the study was to develop a deep learning-based visual and multimodal detection framework for CAD using retinal images. METHODS: We conducted a multicenter cross-sectional study including 383 patients who underwent successful coronary angiography between November 2022 and September 2024 at 4 hospitals. Three models were developed for CAD detection: a convolutional network-based model for retinal images, a hybrid model combining a medical large language model and multilayer perceptron for clinical indicators, and a multimodal model integrating both via a cross-modal attention mechanism. RESULTS: The visual algorithm trained solely on retinal images achieved an area under the receiver operating characteristic curve (AUC) of 0.80 (95% CI: 0.75-0.85), with 90.5% sensitivity and 59.6% specificity. Compared to the CAD consortium clinical score, it showed higher accuracy (76.2%) and sensitivity (87.4%) in the test group. Notably, in the intermediate-risk population (clinical score 15%-85%), it outperformed the clinical indicators-only model with higher AUC (0.79 vs 0.75), accuracy, and sensitivity. Multimodal models combining retinal images and clinical indicators further improved detection, with the best model achieving an AUC of 0.91 (95% CI: 0.88-0.94), 87.0% accuracy, and 92.1% sensitivity. CONCLUSIONS: Our findings suggest that retinal images and vasculature may provide useful information for CAD detection. Retinal imaging offers a noninvasive clinical risk factor and holds promise for developing new diagnostic tools. (Exploring the Application of AI-based Fundus Imaging in Risk Stratification and Risk Assessment Models for Coronary Heart Disease; ChiCTR2400092720).