To create and assess a predictive model for identifying diabetic individuals at heightened risk of diabetic retinopathy (DR) in community health settings using routinely gathered health data.
Approach:
Key Findings:
Prevalence of DR was 13.5%.
GLMNET achieved an AUROC of 0.770 and AUPRC of 0.452 in the validation cohort.
XGBoost showed similar discrimination capabilities, while Ranger performed less favorably.
Urine glucose was identified as the most significant predictor of DR.
Interpretation:
The model demonstrated moderate discrimination and acceptable calibration, indicating potential for improving patient care and risk stratification in primary care settings with limited resources.
Limitations:
The study was conducted in Shenzhen, China, limiting generalizability.
External validation and prospective implementation studies are necessary before clinical application.
Conclusion:
The model's incorporation of routinely gathered variables and clear architecture suggests utility for early risk stratification in primary care, pending further validation through external studies.
