Development and internal validation of an interpretable machine learning model for predicting vancomycin-induced nephrotoxicity in hospitalized children

INTRODUCTION: Vancomycin is widely used for severe Gram‑positive infections in children, but vancomycin‑induced nephrotoxicity (VIN) limits its safe application. Existing prediction models rely primarily on traditional logistic regression with limited discriminative performance in pediatric populations, and machine learning (ML) approaches have not been systematically evaluated. AIM: This study aimed to develop and internally validate an interpretable ML model for predicting VIN in pediatric patients using electronic health record (EHR) data. METHOD: ], preterm birth, blood purification therapy, or major missing data were excluded. VIN was defined as an increase in serum creatinine of ≥ 0.5 mg/dL or > = 50% from baseline on at least two consecutive measurements. Eight ML models were constructed and compared. Model performance was evaluated using the area under the receiver-operating-characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy (ACC), F1 score, Brier score and the area under the precision recall curve (AUPR). The optimal model was interpreted using the Shapley Additive explanation (SHAP) method. Feature sets of 6, 8, 10, and full features were compared using DeLong's test. Clinical utility was assessed via decision curve analysis (DCA). RESULTS: A total of 1,452 children were included, of whom 206 (14.2%) developed VIN. EXtreme gradient boosting (XGboost) achieved the best predictive performance (AUROC = 0.90, AUPR = 0.74, Brier score = 0.07). A simplified 6‑feature XGBoost model (eGFR, amphotericin B, early vancomycin trough concentration, albumin, immunosuppressant use, and acyclovir/ganciclovir co-administration) achieved comparable discrimination to the full model (AUROC = 0.87; P > 0.05 by DeLong's test) with acceptable calibration (ECE 3.8%). DCA suggested favorable net benefit across clinically relevant risk thresholds. CONCLUSION: Our interpretable 6-feature XGBoost model demonstrated promising internal discrimination for pediatric VIN prediction. Pending external validation and temporal evaluation, this model may serve as a hypothesis-generating decision-support framework for risk stratification.