Explainable AI-Based Deep Learning Framework for Predicting Cardiovascular Disease Risk Using Retinal Fundus Images

Shruthi T, Dr. Supriya Shrivastav

Cardiovascular disease (CVD) remains the leading cause of global mortality, necessitating non-invasive, accurate, and interpretable screening tools. Retinal fundus imaging offers an accessible, low-cost means of assessing systemic vascular condition, since microvascular changes visible in the retina tend to parallel those occurring in the coronary vessels. This paper presents a deep learning framework that integrates Contrast-Limited Adaptive Histogram Equalization (CLAHE) preprocessing with a fine-tuned EfficientNet-B0 architecture for binary CVD risk classification. Gradient-weighted Class Activation Mapping (Grad-CAM) is incorporated to generate spatially localized saliency overlays, enabling clinically interpretable predictions by highlighting retinal regions that most influence model decisions. A quadrant-level activation analysis further breaks down attention patterns relative to underlying coronary risk. Evaluation on a publicly available retinal fundus dataset shows that the proposed model attains 94.7% classification accuracy, 93.2% sensitivity, and 95.8% specificity, surpassing several existing deep learning baselines. CLAHE improves vessel contrast and cuts down illumination artifacts, and this turns out to matter a lot for how well the model performs. The Grad-CAM outputs back this up too: the attention patterns line up with what we’d expect clinically, and in high-risk cases we see diffuse activation in the superior-temporal region that matches arteriovenous nicking and vessel tortuosity. Collectively, these findings suggest a practical direction forward: an interpretable retinal imaging pipeline for cardiovascular disease (CVD) screening that could feasibly be integrated into ophthalmology clinics and resource-constrained primary care settings.