Comparative Deep Learning Models for Predicting Diabetic Retinopathy Progression in Saudi Arabia: A Retrospective Observational Study

Abstract Background: Diabetic retinopathy (DR) is a major cause of vision loss amongst adults with diabetes. Early prediction of DR progression could allow timely interventions. We evaluated deep learning models using electronic health records (EHRs) to predict DR progression in Saudi Arabian patients, comparing convolutional neural networks (CNNs), long short-term memory (LSTM) networks and transformer-based architectures. Materials and Methods: We conducted a retrospective cohort study of type 2 diabetes mellitus patients at a tertiary hospital in Eastern Saudi Arabia (2010-2020). Patients had no or mild DR at baseline and at least 3 years of follow-up retinal exams. Demographics, clinical variables and laboratory results were extracted from EHRs. Three deep learning models (1-D CNN, LSTM and Transformer) were trained to predict DR progression, defined as worsening to a more severe DR stage. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC), sensitivity, specificity and other metrics. Risk factors for progression were identified through multivariable logistic regression. Results: Amongst 500 patients (mean age 56.8 ± 12.4 years; 52% male), 110 (22%) experienced DR progression over a median follow-up of 4.5 years. The transformer model had the highest performance (AUC-ROC 0.91, sensitivity 82%, specificity 85%), outperforming the CNN (AUC-ROC 0.86) and LSTM (AUC-ROC 0.88). Logistic regression yielded AUC-ROC 0.80. Key risk factors for progression included baseline HbA1c, diabetes duration, blood pressure and baseline mild DR. Each 1% higher HbA1c increased the odds of progression by 1.3-fold ( P < 0.001). Conclusions: Deep learning models using EHR data effectively predicted DR progression in this Saudi cohort, with the transformer model showing the best performance. Poor glycemic control and longer diabetes duration were strong risk factors for DR worsening.