Model-informed vancomycin precision dosing by population pharmacokinetics combined with machine learning algorithms

Achieving the target AUC/MIC remains a critical challenge in vancomycin therapeutic drug monitoring, with traditional empirical dosing regimens often leading to suboptimal outcomes. This study aimed to develop and validate a novel population pharmacokinetic (PopPK)-informed machine learning model to optimize TDM-guided dose adjustments. By accurately predicting the 24-h area under the curve (AUC), the model assists clinicians in rapidly individualizing regimens, thereby improving target attainment, reducing subsequent modifications and minimizing toxicity.

A total of 1140 vancomycin TDM samples were analysed. Five machine learning models were developed and compared. Model performance assessed using R², RMSE, Lin's concordance correlation coefficient (CCC), mean percentage error (MPE) and mean absolute percentage error (MAPE). The top-performing model (ANN) was subsequently validated on an independent test set. Its performance, interpretability and clinical safety were further examined using SHAP analysis, Bland-Altman plots, error grid analysis (EGA) and ROC curve evaluation.

The ANN (MLP) model demonstrated the best comprehensive performance. It achieved the highest R² (0.920 ± 0.027) and CCC (0.960 ± 0.011), coupled with the lowest RMSE (0.498 ± 0.076 L/h) and minimal systematic bias (MPE: 1.290 ± 1.774). SHAP analysis identified DV (measured concentration), Cys (mg/L) and Last dose as the three most influential predictive features.

The ANN model achieved exceptional accuracy, robustness and clinical safety in predicting individualized vancomycin clearance. This model is deployable as an instantaneous, point-of-care CDSS tool, offering a powerful new pathway to solve the long-standing challenges of accessibility and accuracy in routine TDM practice.