Machine learning and population pharmacokinetics: a hybrid approach for optimizing vancomycin therapy in sepsis patients

Predicting vancomycin exposure is essential for optimizing dosing regimens in sepsis patients. While population pharmacokinetic (PPK) models are commonly used, their performance is limited. Machine learning (ML) models offer advantages over PPK models, but it remains unclear which model—PPK, Bayesian, ML, or hybrid PPK-ML—is best for predicting vancomycin exposure across different clinical scenarios in sepsis patients. This study compares the performance of these models in predicting the 24 hour area under the blood concentration curve (AUC₂₄) to support precision dosing in sepsis care. Data from sepsis patients treated with intravenous vancomycin were sourced from the MIMIC-IV database. The data set was split into training and testing sets, and four models—PPK, Bayesian, ML, and hybrid—were developed. In the testing set, AUC₂₄ was predicted using all models, and performance was evaluated using mean absolute error, mean squared error, root mean squared error, mean absolute percentage error (MAPE), and R^2. A total of 4,059 patients were included. In the absence of vancomycin concentration data, the hybrid model outperformed both PPK and Bayesian models, with MAPE improvements of 58% and 17%, respectively. When vancomycin concentration data were available, the Bayesian model demonstrated the best performance (MAPE: 13.37% vs 68.17%, 34.17%, and 28.52% for PPK, Random Forest, and hybrid models). The hybrid model is recommended to predict AUC₂₄ when concentration data were unavailable, while the Bayesian model should be used when concentrations were available, offering robust strategies for precise vancomycin dosing in sepsis patients.