Prediction of the clinical outcomes of imipenem and meropenem in patients with treatment-associated AKI and death via machine learning models combined with pharmacokinetic parameters: A retrospective cohort study.

ObjectiveThe present study aimed to develop and evaluate machine learning-based predictive models for acute kidney injury (AKI) and all-cause mortality in patients treated with imipenem or meropenem, integrating pharmacokinetic (PK) parameters with routine clinical features.MethodsA total of 235 patients who received imipenem (n = 121) or meropenem (n = 114) between May 2021 and November 2023 were included. Elimination phase and trough concentrations were quantified by UPLC-MS/MS, and a nonlinear mixed-effects model was applied to derive PK/PD parameters. Linear regression was performed to examine associations between baseline characteristics and trough concentrations. Logistic regression was used for initial feature screening, after which 13 machine learning classifiers were developed to predict AKI and mortality. Model discrimination was evaluated using bootstrap internal validation, calibration was assessed via calibration curves, feature importance was examined using SHAP values, and clinical utility was evaluated through decision curve analysis.ResultsThe incidence of AKI was 25.6% (31/121) in the imipenem group and 10.5% (12/114) in the meropenem group. For imipenem-related AKI, the Random Forest classifier yielded the highest discriminative performance (AUC: 0.996, 95% CI: 0.988-1.000); XGBoost performed optimally for meropenem-related AKI (AUC: 0.938, 95% CI: 0.876-1.000). For mortality prediction, Random Forest achieved the best performance for the imipenem group (AUC: 0.963, 95% CI: 0.923-1.000) and XGBoost for the meropenem group (AUC: 0.983, 95% CI: 0.957-1.000). SHAP analysis identified renal function indicators and PK parameters as the predominant contributors to model predictions. Decision curve analysis confirmed favorable net benefit across all optimal models.ConclusionsMachine learning models integrating PK parameters with clinical features demonstrated discriminative ability for predicting carbapenem-associated AKI and mortality. These findings support further investigation into the combination of therapeutic drug monitoring and machine learning analytics for risk stratification, though prospective external validation is required.