Automated Diabetic Retinopathy Severity Grading Using Hybrid Feature Optimization and Weighted Ensemble Learning

Abstract Diabetic Retinopathy (DR) is a leading cause of vision loss in individuals with diabetes. It is caused by damage to the retina’s blood vessels, which can result from chronic hyperglycemia. It is crucial to identify signs of DR early and accurately determine severity before treatment can be initiated, as vision loss can occur. This study focuses on a carefully validated and interpretable hybrid framework for automated DR severity classification instead of introducing a completely new algorithm. The proposed method is a Hybrid Advanced Feature Extraction Framework (HAFEF) and Hybrid Dimensionality Optimisation (HDO) with an optimised weighted ensemble learning strategy. HAFEF extracts spatial, statistical, and frequency-domain features from retinal fundus images after pre-processing, whereas in HDO, the spatial and statistical features of the retinal fundus image are pre-processed. Then, features are selected using PCA and MIF to remove redundancy and improve discriminative ability. This is followed by integrating multiple optimised classifiers, such as random forests, gradient boosting, and support vector machines, into a probability-weighted soft voting ensemble. The framework is tested on a held-out test set and via a nested cross-validation approach, achieving high generalisation and avoiding overfitting (test accuracy: 94.32%, CV accuracy: 93.15%). Experiments on the balanced APTOS 2019 DR data show that the balanced dataset achieves high classification accuracy (94.32%), precision and recall of 0.943, and an ROC-AUC of 0.988. The HAFEF-HDO ensemble framework is a reliable, interpretable and scalable method for automated DR screening and severity grading.