Development and validation of machine learning model for selecting the optimal population pharmacokinetic model for vancomycin.

چکیده اصلی

INTRODUCTION: Therapeutic drug monitoring (TDM) of vancomycin is recommended based on the area under the concentration-time curve (AUC). The Practical Antimicrobial TDM (PAT) software incorporates two population pharmacokinetic (popPK) models developed by Oda et al. and Yasuhara et al.; however, no objective criteria for model selection have been established. This study aimed to develop and validate a machine learning (ML) model to optimize popPK model selection for initial vancomycin dosing. METHODS: The primary outcome was comparison of the two models to identify which yielded a lower absolute relative prediction error for AUC in each patient. Thirty-three clinical variables were screened using Lasso regression to select predictors. Six ML algorithms were compared using nested cross-validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, calibration plots, and decision curve analysis (DCA). RESULTS: Among 531 patients, Oda et al.'s model was more accurate in 60.1% of cases. Lasso selected 22 predictors. Among the algorithms, logistic regression showed the highest mean AUROC (0.720 ± 0.084), whereas the support vector machine demonstrated favorable calibration performance (intercept = 0.01, slope = 0.97). The multi-layer perceptron achieved the lowest Brier score (0.307). Longer hospitalization, lower C-reactive protein (CRP), and female sex were associated with a high predictive probability of accuracy for Oda et al.'s model. DCA suggested the clinical utility of ML-based model selection. CONCLUSION: We developed an ML model for selecting between two popPK models for planning vancomycin dosing regimens in PAT. Considering length of hospital stay, CRP levels, and sex may aid accurate popPK model selection.

متن کامل اصلی