Improving Vancomycin Through Level Prediction with Machine Learning: A Comparative Study of Feature Selection and Model Performance.

Vancomycin is a critical antibiotic used to treat methicillin-resistant Staphylococcus aureus (MRSA) infections, but its narrow therapeutic window necessitates precise dosing to avoid nephrotoxicity and subtherapeutic exposure. Traditional dosing strategies rely on therapeutic drug monitoring (TDM) and Bayesian estimation models like the PAT method, which often exhibit reduced accuracy in critically ill patients, particularly those with impaired renal function. This study explores the use of machine learning techniques to enhance vancomycin trough level prediction, leveraging clinical datasets from the Nippon Medical School, Japan and the open data set MIMIC databases. Three predictive models-Random Forest, LightGBM, and XGBoost-were evaluated based on their mean absolute error (MAE) and feature importance derived from SHAP (Shapley Additive Explanations) analysis. A feature selection approach was implemented to identify key predictors, including creatinine clearance (Ccr), hemoglobin (Hgb), AST, and last-dose timing. Experimental results demonstrate that LightGBM achieved the highest predictive accuracy, outperforming the conventional Bayesian method and providing a robust alternative for individualized dosing.