Developing a Risk Stratification Model Based on Machine Learning for Targeted Screening of Diabetic Retinopathy in the Indian Population

• Published in: Curēus (Palo Alto, CA) CA), 2023-09, Vol.15 (9), p.e45853-e45853
• Main Authors: Surya, Janani, Kashyap, Himanshu, Nadig, Ramya R, Raman, Rajiv
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial intelligence; Blood pressure; Body mass index; Diabetes; Diabetic retinopathy; Epidemiology/Public Health; Fasting; Hemoglobin; Insulin; Machine learning; Neural networks; Ophthalmology; Photography; Population-based studies; Public Health; Risk factors; Statistical analysis; Support vector machines; Variables
• ISSN: 2168-8184; 2168-8184
• DOI: 10.7759/cureus.45853

Objective: This study aimed to develop a predictive risk score model based on deep learning (DL) independent of fundus photography, totally reliant on systemic data through targeted screening from a population-based study to diagnose diabetic retinopathy (DR) in the Indian population.Methods: It involved machine learning application on datasets of a cross-sectional population-based study. A total of 1425 subjects (1175 subjects with known diabetes and 250 with newly diagnosed diabetes) were included in the study. We applied five machine learning algorithms, random forest (RF), logistic regression (LR), support vector machines (SVM), artificial neural networks (ANN), and decision trees (DT), to predict diabetic retinopathy in our datasets. We incorporated a percentage split in the first experiment and randomly divided our data set into 80% as a training set and 20% as a test set. We performed a three-way data split in the second experiment to prevent overestimating predictive performance. We randomly divided our data set into 60% as a training set, 20% as a validation set, and 20% as the test set. Furthermore, we integrated five-fold cross-validation to split the percentage to evaluate our method. We judged the predictive performance based on the receiver operating characteristic (ROC) curve, the area under the curve (AUC), accuracy (Acc), sensitivity, and specificity.Results: The RF classifier achieved the best prediction performance with AUC, Acc, and sensitivity values of 0.91, 0.89, and 0.90, respectively, in the percentage split. Similarly, a three-way data split attained an outcome of 0.86 and 0.85 in AUC and Acc. Likewise, the five-fold cross-validation performed the best with results of 0.90, 0.97, 0.91, and 0.75 in AUC, Acc, sensitivity, and specificity, respectively.Conclusion: Since the RF classifier achieved the best performance, we propose it to identify diabetic retinopathy for targeted screening in the general population.