Randomized trial demonstrates high predictive accuracy for systemic complications in diabetic retinopathy patients, suggesting improved management strategies.
Key Points
This study aims to create a prediction model combining optical coherence tomography angiography and glycated hemoglobin to assess severe systemic complications in diabetic retinopathy.
Retrospective enrollment of 340 type 2 diabetes patients with diabetic retinopathy from January 2020 to December 2024.
Patients were randomly allocated into training (N=238) and validation (N=102) sets at a 7:3 ratio.
Models constructed include logistic regression, gradient boosting machine, and convolutional neural network; performance assessed using AUC and SHAP values.
Multivariate logistic regression identified cardiovascular disease history, diabetes duration, HbA1c, FAZ area, and UACR as independent risk factors (all P < 0.05).
Convolutional neural network achieved AUC of 0.853 in training (95% CI: 0.797–0.909) and 0.820 in validation sets (95% CI: 0.706–0.933), showing superior predictive performance.