Abstract
OBJECTIVE: Although diabetic retinopathy is a leading cause of blindness worldwide, diabetes-related blindness can be prevented through effective screening, detection, and treatment of disease. The study goal was to develop risk stratification algorithms for the onset of retinal complications of diabetes, including proliferative diabetic retinopathy, referable retinopathy, and macular edema. RESEARCH DESIGN AND METHODS: Retrospective cohort analysis of patients from the Kaiser Permanente Northern California Diabetes Registry who had no evidence of diabetic retinopathy at a baseline diabetic retinopathy screening during 2008-2020 was performed. Machine learning and logistic regression prediction models for onset of proliferative diabetic retinopathy, diabetic macular edema, and referable retinopathy detected through routine screening were trained and internally validated. Model performance was assessed using area under the curve (AUC) metrics. RESULTS: The study cohort (N = 276,794) was 51.9% male and 42.1% White. Mean (±SD) age at baseline was 60.0 (±13.1) years. A machine learning XGBoost algorithm was effective in identifying patients who developed proliferative diabetic retinopathy (AUC 0.86; 95% CI, 0.86-0.87), diabetic macular edema (AUC 0.76; 95% CI, 0.75-0.77), and referable retinopathy (AUC 0.78; 95% CI, 0.78-0.79). Similar results were found using a simpler nine-covariate logistic regression model: proliferative diabetic retinopathy (AUC 0.82; 95% CI, 0.80-0.83), diabetic macular edema (AUC 0.73; 95% CI, 0.72-0.74), and referable retinopathy (AUC 0.75; 95% CI, 0.75-0.76). CONCLUSIONS: Relatively simple logistic regression models using nine readily available clinical variables can be used to rank order patients for onset of diabetic eye disease and thereby more efficiently prioritize and target screening for at risk patients.