Development and external validation of machine learning models to predict insulin resistance among iron-deficient children and adolescents.

Background: Insulin resistance (IR) represents a critical metabolic complication in iron-deficient children, yet existing predictive models target general or overweight pediatric populations and do not account for iron-deficiency as a distinct risk modifier. This study aimed to develop and externally validate machine learning (ML) models using routinely available clinical parameters to address this diagnostic gap. Methods: We utilized data from 222 iron-deficient children and adolescents aged 6 to 17 years from the China Health and Nutrition Survey (CHNS) for model training, and 125 cases from two hospitals for external validation. Iron-deficiency was defined using age- and sex-specific soluble transferrin receptor (sTfR) thresholds, with exclusion of elevated high-sensitivity C-reactive protein (hs-CRP) (≥5 mg/L) or missing metabolic variables. IR was defined as Homeostatic Model Assessment for Insulin Resistance (HOMA IR) exceeding 3.0. Least Absolute Shrinkage and Selection Operator (LASSO) regression selected nine predictors from 27 candidate variables (demographics, anthropometrics, blood pressure, hematology, glucose metabolism, lipids, hepatic and renal function). Four ML algorithms [logistic regression (LR), random forest (RF), k-nearest neighbor (KNN), and extreme gradient boosting (XGBoost)] were developed and evaluated by area under the curve, sensitivity, specificity, and calibration, with five-fold repeated cross-validation for internal validation. SHapley Additive exPlanations (SHAP) analysis quantified predictor contributions. Results: XGBoost achieved optimal discriminative performance with an external validation area under the receiver operating characteristic curve (AUC) of 0.940 [95% confidence interval (CI): 0.889-0.991], outperforming other algorithms. RF demonstrated the highest training AUC (0.993, 95% CI: 0.987-1.000) with near-perfect sensitivity (0.985, 95% CI: 0.920-1.000) but showed limited generalization capacity given minimal training-validation divergence. LR and KNN achieved lower validation AUC values of 0.832 (95% CI: 0.743-0.922) and 0.823 (95% CI: 0.740-0.905), respectively. XGBoost was selected as the final model based on superior specificity (0.967, 95% CI: 0.906-0.993) and tighter CIs, indicating more stable performance estimation. Fasting glucose (mean |SHAP| =0.707) and triglycerides (0.383) emerged as dominant predictors, while albumin demonstrated a protective association [odds ratio (OR) 0.86, 95% CI: 0.78-0.95]. Conclusions: This study establishes an externally validated, interpretable ML framework for predicting IR among iron-deficient youth using routine clinical data. While the XGBoost model demonstrates promising discriminative performance and geographic generalizability, the modest sample size and single-province validation limit immediate deployment readiness. Prospective multi-site validation is required before any consideration of clinical implementation as a developmental screening framework.