<p>Diabetic retinopathy (DR) is a severe complication of diabetes that damages retinal blood vessels, potentially causing irreversible blindness. Existing deep learning (DL) models for DR detection are often limited to binary classification or single convolutional neural network (CNN) architectures, reducing their ability to generalize across diverse datasets. Moreover, challenges such as class imbalance and high inter-class similarity hinder accurate multistage classification. To overcome these issues, this study proposes a hierarchical deep learning ensemble model for enhanced DR identification and multistage classification. The framework integrates nine different CNNs via transfer learning, systematically combining ensemble strategies within a two-level hierarchical architecture to exploit complementary model strengths. Four ensemble techniques—simple averaging, weighted averaging, stacking, and majority voting—were integrated into two-level ensemble blocks, with outputs combined at the final stage for classification. Experiments were conducted on the APTOS 2019 and IDRiD datasets utilized preprocessing techniques such as resizing, Gaussian blur, Ben Graham enhancement, cropping, and data augmentation to improve robustness. To address class imbalance, minority DR classes were oversampled and augmented to ensure balanced representation across disease stages. Among all methods, the stacking ensemble achieved the highest performance with 97.4% accuracy, outperforming individual CNNs and demonstrating superior precision, recall, and F1-scores. The system also generates a structured DR severity report indicating the predicted stage and confidence level, supporting ophthalmologists to efficiently assess disease progression and prioritize patient management. Future work will refine the stacking ensemble through adaptive weighting and expand validation using larger, multi-institutional datasets to enhance robustness and clinical applicability.</p>