Integrated Decision Support System for Diabetic Retinopathy Grading Using Explainable AI and Uncertainty Quantification

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, yet screening coverage remains critically low due to a shortage of trained ophthalmologists and other primary eye care professionals [1] [2]. This article presents an Integrated Decision Support System (IDSS) for ocular health that employs a fine-tuned EfficientNet-B3 convolutional neural network, trained on the APTOS 2019 Blindness Detection dataset (3,662 fundus images), to perform five-class DR grading. The model achieved a maximum area under the receiver operating characteristic curve (AUC) of 0.9858 and a maximum Quadratic Weighted Kappa (QWK) of 0.8789 after 30 training epochs. The system incorporates Gradient-weighted Class Activation Mapping++ (Grad-CAM++) for spatial attribution and Monte Carlo Dropout for predictive uncertainty quantification, integrated into a Streamlit clinical interface with structured risk stratification. A multi-task extension simultaneously learns DR grading and glaucoma risk scoring, achieving AUC values of 0.9841 and 0.9849, respectively. The complete system demonstrates that a functional, explainable AI prototype can be developed within a one-month engineering timeline using open-source tools. Keywords — Clinical Decision Support, Convolutional Neural Network, Diabetic Retinopathy, Explainable Artificial Intelligence, Machine Learning, Uncertainty Quantification.