An Explainable AI-Based Predictive Modeling Framework for Early Detection of Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is the leading preventable cause of blindness in working-age adults, yet its asymptomatic early stages result in widespread diagnostic delay in resource-constrained healthcare settings. This study presents an Explainable Artificial Intelligence (XAI) framework designated RetineXAI that integrates a hybrid deep learning architecture with model-agnostic and model-specific interpretability modules to enable clinically transparent, automated five-stage DR grading from fundus photographs. RetineXAI combines a dual-branch convolutional neural network (EfficientNetV2-L encoder fused with a Swin Transformer branch) with an ensemble decision layer and post-hoc explainability via Gradient-weighted Class Activation Mapping (Grad-CAM++), SHapley Additive exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME). The model was trained on 88,702 fundus images from four benchmark datasets (EyePACS, APTOS-2019, Messidor-2, IDRiD) after a rigorous multi-stage preprocessing pipeline including adaptive histogram equalization (CLAHE), optic-disc normalisation, and lesion-aware augmentation. On held-out external validation, RetineXAI achieved an overall accuracy of 96.14%, macro-averaged AUC-ROC of 0.9871, sensitivity of 95.82%, specificity of 97.43%, and F1-score of 0.9601 across five DR severity grades. For clinically critical early-stage (mild NPDR) detection, sensitivity reached 94.37% with AUC 0.9812, outperforming six state-of-the-art baselines. Grad-CAM++ saliency maps localised pathology-consistent lesion signatures (microaneurysms, hard exudates, neovascularisation) with a mean Intersection-over-Union (mIoU) of 0.831 against expert-annotated lesion masks. Ophthalmologist-in-the-loop evaluation with 12 retinal specialists confirmed that XAI-augmented clinical decision support improved diagnostic accuracy by 11.4 percentage points and reduced mean grading time from 6.3 to 3.1 minutes per image. RetineXAI demonstrates that high-accuracy, interpretable AI can be deployed as a first-line screening tool in teleophthalmology and community screening programmes.