Development and validation of machine learning models to predict vancomycin- and teicoplanin-associated acute kidney...

Despite the effectiveness of vancomycin and teicoplanin in managing Gram-positive infections, their nephrotoxicity may prolong hospitalization, and increase morbidity, mortality, and healthcare costs. This study aimed to develop and validate clinically applicable prognostic machine learning models for predicting vancomycin-associated acute kidney injury (VA-AKI) and teicoplanin-associated AKI (TA-AKI).

This retrospective study in Taiwan utilized the Taipei Medical University Clinical Research Database. Patients receiving intravenous vancomycin or teicoplanin therapy between February 2010 and December 2020 were included. Features were selected from 198 variables through recursive feature elimination using feature importance (RFECV) and SHapley Additive exPlanations importance (ShapRFECV). Twelve models were constructed using XGBoost and LightGBM. Model performance was assessed by eight evaluation metrics, including the area under the receiver operating characteristic curve (AUROC). The optimal threshold was determined based on the maximum F₁ score, and the SHAP analysis assisted in model interpretation.

Among 9342 included patients, 19.70% (1383/7020) of patients in the training set, 18.58% (326/1755) in the internal validation set, and 20.5% (116/567) in the external validation set developed AKI. The XGBoost model, using features selected by ShapRFECV, demonstrated optimal predictive performance in both internal (AUROC 0.798, 95% confidence interval [CI] 0.791-0.804) and external (AUROC 0.779, 95% CI: 0.767-0.791) validation.

The XGBoost model, leveraging time-series data, accurately predicts AKI in vancomycin and teicoplanin users, supporting early risk assessment and personalized patient management. Future multicentre prospective external validation is needed to strengthen its real-world applicability and generalizability.