AI-Driven Retinal Screening: A Stacked Machine Learning Framework for Automated Diabetic Retinopathy Detection in Routine Medical Checkups Using a Public Fundus Database

Diabetic retinopathy (DR) is a leading cause of preventable blindness, yet large numbers of people with diabetes miss their annual eye examination because grading retinal photographs requires scarce specialists. This paper proposes an artificial intelligence (AI) based eye-inspection framework that embeds automated DR screening inside a routine general medical checkup. The framework comprises five layers: image acquisition, automated quality control, lesion feature extraction, a stacked machine learning (ML) decision engine, and an explainable clinical report with referral. We evaluate the approach on the public UCI Diabetic Retinopathy Debrecen database, which contains 19 lesion, anatomical and image-level features extracted from 1,151 MESSIDOR fundus images. Nine classical and neural classifiers are benchmarked with 10-fold stratified cross-validation, after which the proposed pipeline adds lesion-aware feature engineering, hyperparameter tuning and stacked generalization. The proposed model reaches a cross-validated area under the ROC curve (AUC) of 0.854 and accuracy of 77.6%, improving the logistic-regression baseline by 0.029 AUC, and achieves a hold-out AUC of 0.878 (95% bootstrap CI 0.830–0.917) with 79.7% sensitivity and 77.6% specificity. Lowering the decision threshold to 0.30 yields 91.9% sensitivity for screening use. In an extension to cataract, three image-clarity features separate normal from cataract fundus images with a cross-validated AUC of 0.978. Permutation importance and ablation show that microaneurysm counts carry most of the diagnostic signal, consistent with clinical grading. Using CDC surveillance data covering 9.6 million Americans with DR, we project that the screening operating point would detect 8.8 million cases while removing about 16 million specialist grading visits per year. We discuss clinical workflow integration, challenges including domain shift, bias and regulation, and a roadmap towards deep-learning-based multi-disease retinal screening.