Development and Validation of a Diabetic Retinopathy Risk Stratification Algorithm.

Area of Science:

• Ophthalmology
• Data Science
• Public Health

Background:

• Diabetic retinopathy is a leading cause of preventable blindness globally.
• Early detection and treatment are crucial for managing diabetes-related eye complications.

Purpose of the Study:

• To develop risk stratification algorithms for predicting diabetic retinopathy onset.
• To identify patients at high risk for proliferative diabetic retinopathy, referable retinopathy, and macular edema.

Main Methods:

• Retrospective cohort analysis of 276,974 patients from the Kaiser Permanente Northern California Diabetes Registry (2008-2020).
• Development and internal validation of machine learning (XGBoost) and logistic regression models.
• Performance assessment using Area Under the Curve (AUC) metrics.

Main Results:

• The XGBoost model achieved AUCs of 0.86 for proliferative diabetic retinopathy, 0.76 for diabetic macular edema, and 0.78 for referable retinopathy.
• A simpler nine-covariate logistic regression model showed comparable performance (AUCs 0.82, 0.73, and 0.75, respectively).

Conclusions:

• Simple logistic regression models using nine clinical variables can effectively rank patients by diabetic eye disease risk.
• These models enable more efficient prioritization and targeting of screening for at-risk individuals.
• Improved screening strategies can help prevent diabetes-related blindness.