Decoding diabetic retinopathy risk through hybrid modeling: integrative assessment of machine learning algorithms and...

Diabetic retinopathy (DR) is a major microvascular complication of diabetes mellitus and a leading cause of irreversible visual loss. Insulin resistance plays a central role in DR pathogenesis, but direct measurement is challenging in large-scale studies. The estimated glucose disposal rate (eGDR) is a validated surrogate of insulin sensitivity, yet its relationship with DR in population-based cohorts remains underexplored.

Data from 954 adults with diabetes in the 2015-2020 National Health and Nutrition Examination Survey were analyzed. Sociodemographic, behavioral, and cardiometabolic factors were compared between groups, followed by multivariable logistic regression, restricted cubic spline (RCS) modeling, and threshold effect analyses. Machine learning algorithms were applied for predictive modeling and feature ranking.

Each one-unit increase in eGDR was associated with 21% lower odds of DR (OR = 0.79; p < 0.001). Quartile analysis revealed a graded inverse association, most pronounced in the third quartile (OR = 0.53; p = 0.020), with a significant linear trend (p < 0.001). RCS confirmed a robust association with non-linearity (p = 0.029) and a plateau at higher eGDR values. Two-piecewise regression identified an inflection at eGDR = 3.92, below which the association weakened (p = 0.010). Among eight algorithms, XGBoost achieved the highest discrimination (AUC = 0.894), with Shapley Additive Explanations ranking eGDR among the most influential predictors.

These findings support eGDR as an independent, non-linear, and potentially threshold-dependent biomarker for DR risk stratification, with clinical implications for targeted screening and intervention.

The online version contains supplementary material available at 10.1007/s40200-025-01765-8.