End-to-End Development of a Multimodal Fusion Model for Diabetic Retinopathy and Cardiovascular Risk Prediction

: Diabetic Retinopathy (DR) and Cardiovascular Disease (CVD) are serious consequences of diabetes, yet current diagnostic processes address them independently. This research describes a compact end-to-end multi-modal approach for predicting DR severity and a continuous cardiovascular risk (CVR) score by combining retinal fundus pictures with structured clinical data. The system consists of an EfficientNet-B3 retinal encoder, an autoencoder-based clinical encoder, and a Transformer fusion module that monitors relationships between ocular biomarkers and systemic health indicators. To overcome class imbalance and the scarcity of high-risk data, Conditional GANs and Tabular VAEs create realistic synthetic pictures and clinical records. Experiments on the BRSET dataset demonstrate that the suggested technique outperforms unimodal and classic fusion base-lines, with an AUC of 0.95 and F1 of 0.90 for DR classification and an MSE of 0.021 for CVR regression. These findings highlight the potential of multimodal learning for unified, early detection of diabetes-related ocular and cardiovascular problems.