Retinal vascular biomarker concept learning with multimodal fusion for cardiovascular disease detection

Cardiovascular disease (CVD) remains a leading cause of death and disability worldwide. Although coronary angiography is regarded as the clinical gold standard for CVD diagnosis, its invasiveness, high cost, and limited patient acceptance hinder timely screening and early detection. Retinal fundus imaging provides a non-invasive, low-cost, and repeatable method to assess systemic microvascular health. However, existing methods do not adequately capture the intrinsic association between retinal vascular morphology and CVD. To address this issue, we propose a deep learning framework for non-invasive CVD detection that integrates retinal vascular biomarkers and clinical indicators via concept learning and multimodal fusion. The model includes three modules: (1) a biomarker concept perception module (BCPM), which constructs learnable concept tokens to quantify vascular morphological variations; (2) a clinical feature representation module (CFRM), which employs a dual-path encoding strategy to model categorical and continuous clinical variables separately; and (3) a cross-modal attention aggregation module (CAAM), which uses disease tokens as queries to fuse retinal and clinical features for CVD diagnosis. Across the United Kingdom Biobank (UK Biobank) and a private dataset, the proposed model achieves area under the receiver operating characteristic curve (AUC) values of 0.8524 ± 0.0164 and 0.7930 ± 0.0147, area under the precision–recall curve (AUPR) values of 0.6576 ± 0.0393 and 0.6171 ± 0.0326, and F1 scores of 0.6190 ± 0.0207 and 0.5887 ± 0.0124, respectively, outperforming existing methods. Further analysis shows that decreased vessel density and vein density are significantly associated with higher CVD risk, with stronger associations in specific subgroups.