Development and validation of a machine learning model for individualised meropenem dosing in CRRT patients.

Optimizing meropenem dosing in critically ill patients undergoing continuous renal replacement therapy (CRRT) remains challenging due to significant pharmacokinetic variability. This study aimed to develop and validate a machine learning (ML)-based clinical decision support system (CDSS) for individualizing meropenem dosing in this population. A large-scale virtual cohort (n = 48,000) was generated using a published population pharmacokinetic (POPPK) model to simulate diverse dosing regimens. Various ML models were trained to predict the probability of achieving two stringent pharmacodynamic (PD) targets: 100% fT>MIC and 100% fT>4×MIC. Feature importance was interpreted using SHAP (Shapley Additive exPlanations) analysis. External validation was conducted using clinical CRRT patient data, and predictive performance was compared with that of the traditional POPPK model. The top-performing model was deployed into a Streamlit-based web CDSS, named "MerDose" for clinical use. The ML models exhibited excellent predictive performance on the simulated test set, with accuracy and F1 scores predominantly above 0.95. Notably, in the external validation cohort, the Gradient Boosting (GB) model robustly outperformed the conventional POPPK model across stringent pharmacodynamic targets (100% fT>MIC and 100% fT>4×MIC), achieving AUC values of 0.884 and 0.929, respectively, versus 0.823 and 0.871 for the POPPK model. Interpretation via SHAP analysis confirmed that the minimum inhibitory concentration (MIC), dosing interval, dose amount, CRRT clearance (CL