Improving Diabetic Retinopathy Prediction with Stacked Generalization and Out-of-Fold Learning

Abstract

Diabetic retinopathy (DR) is a serious microvascular complication of diabetes that causes significant
vision impairment worldwide. Because DR is usually asymptomatic in its early stages, frequent retinal
screening is essential to prevent permanent blindness. However, manual screening is time-consuming,
labour-intensive, and difficult to scale in regions with limited expertise. These challenges have led to the
development of automated computer-aided systems that enable effective, inexpensive monitoring of
retinal disease.
This paper, motivated by advances in ensemble learning for medical decision support, proposes a
stacking-based classification system to predict retinal abnormalities using the Retinadataset. The
architecture consists of four base learners—Random Forest (RF), Extra Trees(ET), Histogram Gradient
Boosting, and Logistic Regression(LR)—whose out-of-fold predictions, produced via 10-fold stratified
cross-validation, are combined to form a meta-level feature matrix. AnLR meta-learner generates the
final predictions.
The proposed system was evaluated against various single-base models, including LR,RF, Support
Vector Machines(SVM), Naive Bayes(NB), Decision Trees(DT), and a blending ensemble.
Experimental results demonstrate that the ensemble approach is more predictive, robust, and balanced
in classification than individual learners. The Tuned Ensemble achieved the highest performance:
77.15% accuracy, 70.87% sensitivity, 84.26% specificity, 83.59% precision, 76.71% F1-score, 54.58%
Kappa, and an 84.09% area under the ROC curve (AUC). The Blending Ensemble achieved the best
AUC and a slightly higher sensitivity, while LR was the best-performing single classifier.
These findings suggest that stacking-based ensemble learning provides a strong, interpretable framework
for the automated screening of retinal abnormalities and could supplement clinical risk assessment in the
early stages.