Explainable retinal deep learning for cardiovascular risk stratification: a multiple modality analysis framework with vascular-centric interpretability and robustness

Early non-invasive prediction of cardiovascular risk is an essential prerequisite for preventive medicine, especially in resource-limited settings. Retinal imaging provides unique insights into systemic vascular health; however, most existing deep learning models lack pathophysiological explainability and generalizability, which limits their clinical adoption. Furthermore, existing approaches fail to effectively integrate specific retinal biomarkers with multimodal indicators of cardiovascular conditions and are not adaptable to variations in demographics or imaging conditions. This limits their applicability in real-time environments, particularly within SDN-enabled 5G healthcare infrastructures, where explainability and reliability are essential. In this work, a comprehensive and explainable deep learning architecture is developed for retinal-based cardiovascular risk assessment. The proposed framework integrates five analytical components: (1) MARGE-Net, which incorporates clinical biomarkers with vascular topology; (2) SCCIM, which visualizes causally important retinal features; (3) HVAAT, a hierarchical transformer for modelling vessel attributes; (4) PaVSCM, which ensures anatomical plausibility across populations; and (5) ReCVD-LDM, which separates imaging artifacts from true anatomical predictors using contrastive learning. These modules collectively improve accuracy, interpretability, and robustness, while enabling compatibility with real-time deployment in Software-Defined Networking (SDN)-enabled 5G healthcare infrastructures. This is further supported by a conceptual edge-cloud architecture that enables efficient data transmission and low-latency inference. The model is validated on publicly available datasets, including MESSIDOR, AV-WIDE, and the STARE VAS subset, achieving an AUC of 0.94, with a 31% improvement in interpretability and 26% reduction in prediction errors caused by imaging artifacts. Overall, this work advances trustworthy and interpretable AI for cardiovascular risk prediction from retinal images in next-generation healthcare systems.