An interpretable XGBoost model for predicting measured teicoplanin plasma concentrations in ICU patients with bloodstream infection: a two-center real-world validation study

Background Profound pharmacokinetic variability complicates individualized teicoplanin therapy in critically ill patients with bloodstream infection. We aimed to develop and externally validate an interpretable machine learning (ML) model for predicting teicoplanin plasma concentrations in this population. Methods This retrospective two-center real-world study included 260 intensive care unit patients from one center as the modeling cohort (randomly split 8:2 into training and internal test sets) and 91 patients from an independent center as the external validation cohort. Ten ML and deep learning algorithms were constructed using demographic, clinical, and laboratory variables. The optimal model was interpreted via Shapley additive explanations (SHAP). Results The extreme gradient boosting (XGBoost) model achieved the most favorable point estimates. On the internal test set (n = 52): the model yielded an R 2 of 0.73 (95% CI 0.56–0.81), RMSE of 3.22 (95% CI 2.62–3.77), MAE of 2.52 (95% CI 1.98–3.10), and ±30% prediction accuracy (P30%) of 78.8% (95% CI 67.3–88.5). On the external validation cohort (n = 91), it achieved an R 2 of 0.76 (95% CI 0.69–0.80), RMSE of 3.61 (95% CI 3.08–4.14), MAE of 2.89 (95% CI 2.48–3.35), and P30% of 69.2% (95% CI 59.3–78.0). After Holm–Bonferroni correction, paired bootstrap comparisons showed XGBoost significantly outperformed 5 of 9 comparator algorithms in the internal test set and 7 of 9 in the external cohort, with performance comparable to several other tree-based ensemble methods. SHAP analysis identified daily dose, hemodialysis, and mechanical ventilation as features positively associated with predicted teicoplanin concentrations. Conclusion This internally and externally validated interpretable XGBoost model delivers consistent cross-cohort performance, and serves as a dose-anchored adjunct to conventional therapeutic drug monitoring, providing a preliminary methodological foundation for exploring individualized teicoplanin dosing strategies.