Improvement of Risk Stratification for Diabetic Retinopathy Progression in Primary Care via an AI Model with Dynamic Anthropometric Data

Aim: To identify predictors of diabetic retinopathy (DR) progression on the basis of baseline data and annual dynamic changes and to develop an AI-driven personalized risk prediction model for improving clinical referrals. Methods: In this retrospective case‒control study, fundus images from 601 participants (273 progressors and 328 nonprogressors) were quantitatively analyzed via an AI system. The participants were categorized on the basis of second-year fundus changes. After random data splitting (70% training, 30% validation), LASSO and multivariate logistic regression were used for feature selection and model construction. Model performance was assessed by the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Results: The dominant risk factors were modifiable variables: elevated HbA1c (OR = 1.473, 95% CI: 1.249–1.739), increased BMI (OR = 1.457, 95% CI: 1.214–1.747) and increased waist circumference (OR = 1.083, 95% CI: 1.039–1.128). The model demonstrated robust and consistent discrimination, with AUCs of 0.799 (training) and 0.788 (validation). DCA showed clinical utility across a wide threshold probability range (0--0.85). Conclusions: Proactive management of modifiable risk factors is crucial. The developed AI-driven personalized prediction model shows robust performance and holds promise for enhancing referral precision in primary care, offering a practical strategy to improve DR management.