Risk factor identification and classification of diabetic retinopathy among Northeast Indian population using machine learning models

Problem considered

Diabetic retinopathy (DR) is the most common cause of blindness among working-age population. With escalating global diabetes prevalence, identifying risk factors is crucial for prioritizing DR diagnosis. This study aimed to determine key risk factors in the Northeast Indian population and classify DR using artificial intelligence models.

Methods

In this study, twenty-seven clinical and biochemical characteristics of 188 individuals across four groups, healthy control (HC), type-2 diabetes mellitus (T2DM), non-proliferative DR (NPDR), and proliferative DR (PDR) were analysed to identify the DR risk factors. Data were analysed using four ML models, and Shapley Additive Explanation (SHAP) analysis was applied to the best-performing random forest (RF) model to assess the clinical relevance of features. Additionally, convolutional neural network (CNN) and graph neural network (GNN) models were employed for DR classification using fundus images.

Results

Among four classifiers, the RF model achieved 100 % training accuracy and 92 % test accuracy. In the testing dataset, the RF model achieved an area under the ROC curve (AUC) of 1 for HC and PDR. While it achieved an AUC of 0.96 and 0.97 for T2DM and NPDR, respectively. SHAP analysis identified uric acid levels, T2DM duration, glycosylated haemoglobin (HbA1c), fasting blood sugar (FBS), and tobacco/betelnut chewing as significant predictors of DR. The GNN model outperformed CNN in fundus image classification, achieving 82 % test accuracy and an AUC of 0.85.

Conclusion

The RF model effectively identified DR risk factors in the Northeast Indian population, while GNN demonstrated robust classification accuracy. Integrating ML and DL enhances early DR risk assessment and diagnosis, improving disease management and patient outcomes.