People with diabetes can develop diabetic retinopathy (DR), an eye disease that is a severe issue that affects more than a quarter of people with diabetes. The importance of early DR identification has increased since diabetes prevalence, particularly in developing countries, is on the rise. The prevention of DR begins with early detection and treatment. Detecting the DR condition is a time-consuming and labor-intensive process requiring an expert to analyze the images of the patient’s retina. It makes slow up in diagnosis and treatment. However, automated examination of retinal images could significantly increase the effectiveness and coverage of DR screening programs. In light of this, we propose two predictive models—transfer learning and CNN—that incorporate the Alex model. The study shows that deep learning has great potential, and the EfficientNet transfer learning model achieves an impressive 0.80 accuracy with excellent recall, precision, and F-score metrics. While achieving a respectable 0.7347 accuracy, the CNN with Alex variation must be higher than its counterpart. The experimental outcomes conclusively demonstrate that the proposed approach surpasses traditional methods in utility and effectiveness. Our study enhances the field by revealing the significance of deep learning and transfer learning in DR severity prediction, ushering in a more effective era of diagnosis.