Development of a risk stratification tool for rapidly progressive diabetic retinopathy in type 2 diabetes.

BACKGROUND

The progression of rapidly progressive diabetic retinopathy (PDR) in type 2 diabetes mellitus (T2DM) is characterized by substantial inter-individual variability. To develop and validate a nomogram for individualized risk prediction and stratification of rapidly progressive PDR in T2DM by incorporating diabetes duration, glycated hemoglobin (HbA1c), 24-hour urinary protein quantification, growth differentiation factor 15 (GDF15), Diabetic Retinopathy Severity Scale (DRSS) grade, and foveal avascular zone area.

METHODS

This retrospective study enrolled 342 patients with T2DM (1999 WHO criteria), randomly assigned to training (n=240) and validation (n=102) sets (7:3 ratio). Baseline demographic, clinical, metabolic, renal, inflammatory biomarker, and ophthalmic imaging data were collected. Predictive variables were selected via univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Independent predictors identified by multivariable logistic regression were incorporated into a nomogram. For comparison, Random Forest, multivariable logistic regression, and Gradient Boosting Machine models were also developed. Model performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC), calibration curves, and decision curve analysis (DCA).

RESULTS

Univariate analysis identified six significant factors (all P < 0.05): diabetes duration, HbA1c, 24-hour urinary protein quantification, GDF15, DRSS grade, and foveal avascular zone area. LASSO regression retained all six, and multivariable logistic regression confirmed them as independent risk factors for rapidly progressive DR in T2DM (all P < 0.05). Three machine learning models were constructed. The Random Forest model achieved the numerically highest validation AUC (0.780) compared with Gradient Boosting Machine (0.741) and multivariable logistic regression (0.698), though the DeLong test showed no statistically significant difference between Random Forest and Gradient Boosting Machine (P = 0.38). Calibration curves showed good consistency between predicted and observed probabilities. DCA indicated high clinical net benefit of the model at 0.1-0.8 threshold probability vs other models and extreme strategies.

CONCLUSION

A novel risk prediction model for rapidly progressive PDR in T2DM was developed and validated by integrating multidimensional parameters. Demonstrating favorable discrimination, calibration, and clinical utility, this model provides a promising tool for early identification of high-risk individuals and optimization of personalized intervention strategies.