Comparing Insulin Resistance Surrogates for Prevalent Kidney Disease in CHARLS: An Explainable Machine Learning Study
Wei Wang, Qin LuBackground
Insulin resistance (IR) is related to renal injury, but the relative contribution of routinely calculated IR surrogates to classifying prevalent kidney disease remains uncertain. We compared a broad panel of these indices using an interpretable machine learning framework.
Methods
This cross‐sectional analysis included 9644 participants from Wave 3 (2015) of the China Health and Retirement Longitudinal Study (CHARLS). Twelve IR‐related surrogate indices were calculated, including TyG, TyG‐BMI, TyG‐WC, and METS‐IR. Ten classification algorithms were developed using a 70% training subset and a 30% internal hold‐out test subset. Model performance was described using AUROC together with calibration and decision‐curve analysis, and SHAP was used to describe model‐generated feature contributions.
Results
In the internal hold‐out subset, GBM had the numerically highest AUROC (0.6800), followed by CatBoost (0.6611), logistic regression (0.6597), and LASSO (0.6596); these values indicate modest discrimination. GBM and CatBoost were subsequently illustrated with SHAP. Among the 12 IR surrogates, TyG‐WC, METS‐IR, TyG‐ABSI, and TyG showed relatively large model‐attribution values.
Conclusion
In this cross‐sectional CHARLS analysis, routinely derived IR surrogates contributed information to the classification of prevalent self‐reported kidney disease. SHAP results describe how variables contributed to the fitted models.