Development and validation of a machine learning–based early warning model for carbapenem-resistant Klebsiella pneumoniae bloodstream infections using non-carbapenem susceptibility profiles

bloodstream isolates collected from 60 hospitals in China between 2014 and 2023. Non-carbapenem antimicrobial susceptibility interpretations were used as model inputs, while carbapenem results were excluded. Data were split chronologically into training (2014-2021), validation (2022), and test (2023) sets. Logistic regression, XGBoost, and CatBoost models were developed and evaluated using discrimination, calibration, decision curve analysis (DCA), and SHAP-based interpretability. XGBoost demonstrated the best overall performance, achieving higher discrimination with a ROC-AUC of 0.993 and a PR-AUC of 0.973 on the test set, along with superior calibration as reflected by the lowest Brier score (0.018). At a sensitivity-targeted threshold (~0.95), XGBoost maintained high sensitivity (0.924), excellent specificity (0.989), and a favorable positive predictive value (0.944), while preserving a high negative predictive value (0.986). SHAP analysis identified key non-carbapenem susceptibility features contributing to CRKP risk prediction. Non-carbapenem susceptibility profiles enable early identification of CRKP bloodstream infections. A machine learning-based early warning model, particularly XGBoost, may support laboratory-based risk stratification and complement conventional susceptibility testing.