A Machine Learning-Based Model for Individualized Prediction of Vancomycin Concentration-Time Curves in ICU Patients

Therapeutic drug monitoring is essential for ensuring the efficacy and safety of vancomycin therapy in critically ill patients. This study aimed to develop a machine learning model for individualized prediction of vancomycin concentration-time curves in ICU patients. Adult ICU patients who received intravenous vancomycin and underwent therapeutic drug monitoring at Peking Union Medical College Hospital between January 2014 and December 2023 were retrospectively included. A total of 401 patients were randomly divided into training (n = 280) and testing (n = 121) cohorts. Individual pharmacokinetic parameters were estimated using Bayesian posterior inference and served as reference targets. Five machine learning algorithms were evaluated, and the two with the best predictive performance, Lasso Regression and LightGBM, were integrated with a one-compartment pharmacokinetic model to construct the final predictive model. In the internal testing cohort, the model achieved a mean absolute percentage error (MAPE) of 39.5% for vancomycin concentration prediction. External validation in an independent cohort of 2283 patients showed consistent performance (MAPE = 35.6%). The machine learning-based model significantly outperformed the classic pharmacokinetic model (p < 0.001) in both internal and external validations. A user-friendly software tool based on the model was also developed to facilitate clinical implementation. These findings suggest that the proposed model offers a robust and practical decision-support tool for optimizing individualized vancomycin dosing in ICU settings. Trial Registration: ClinicalTrials.gov identifier: NCT06431412.