RetinaFusion-Net: An Explainable Multimodal Deep Learning Framework for Shortest-Path Multi-Organ Clinical Risk Prediction

Abstract

Retinal disorders are one of the most important causes of irreversible loss of vision and are emerging as early markers of systemic disorders of the cardiovascular, neurological and renal systems. Current diagnostic systems, however, are mostly based on isolated retinal abnormalities and do not model disease risk propagation between various organs nor do they include environmental/experimental or clinical influences. To overcome this, this study is proposing a multimodal deep learning model based explainable framework RetinaFusion-Net and a Shortest-Path Risk Propagation (SPRP) model for multi-organ clinical risk prediction. The framework is a hybrid fusion model of CNN and Vision Transformer that integrates fundus images, optical coherence tomography (OCT), clinical parameters and environmental factors. Physiological relationships are represented by a weighted organ graph and the SPRP algorithm recognizes the most likely disease progression pathway from retinal biomarkers to vulnerable organs. Explainable AI techniques are used to enhance the clinical interpretability of AI systems by augmenting the patient's retina with indications of influential features and predicted risk pathways. Experimental evaluation on multimodal retinal datasets shows an accuracy of 98.74%, a precision of 98.41%, a recall of 98.58%, an F1 score of 98.49% and an AUC of 99.12%, which improves the results obtained by the current deep learning techniques. Through intelligent analysis of retinal biomarkers, the proposed framework has the capability of early detection of retinal disease, accurate clinical risk prediction for multiple organs, as well as interpretative and intelligent decision support, offering a scalable approach to precision ophthalmology and preventive health care.