Exposure prediction and dose optimization of polymyxin B based on bayesian and machine learning

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Computational Biology and Bioinformatics
• Clinical Pharmacology

Background:

• Polymyxin B (PMB) is a critical antibiotic, but its narrow therapeutic index necessitates precise exposure monitoring.
• Predicting PMB exposure is challenging due to complex pharmacokinetic variability.
• Accurate prediction of drug exposure aids in optimizing dosing regimens and improving patient outcomes.

Purpose of the Study:

• To evaluate the application scenarios of maximum a posteriori Bayesian estimation (MAP-BE) and eXtreme Gradient Boosting (XGBoost) for predicting PMB exposure.
• To compare the predictive performance of MAP-BE and XGBoost in various simulation scenarios.
• To identify optimal sampling strategies for accurate PMB pharmacokinetic modeling.

Main Methods:

• Developed and tested MAP-BE and XGBoost models using two sets of simulations based on a PMB population pharmacokinetic (PopPK) model.
• Evaluated predictive performance using metrics such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).
• Assessed model accuracy under different sampling strategies, including dense point and single-trough concentrations.

Main Results:

• Both MAP-BE and XGBoost accurately estimated the area under the concentration-time curve (AUC0-12h) under a dense point sampling strategy.
• A single 6-h sampling strategy yielded the best prediction accuracy with negligible bias (RMSE < 1 mg·h/L, MAE < 1 mg·h/L, R² > 0.99).
• XGBoost outperformed MAP-BE in accuracy and efficiency for fitting single-trough concentrations, with its 12-h model capturing temporal fluctuations.

Conclusions:

• MAP-BE and XGBoost are viable methods for predicting PMB AUC, aiding in the selection of optimal prediction approaches.
• The findings support the use of therapeutic drug monitoring data to guide PMB dosing adjustments.
• XGBoost shows promise as a more accurate and efficient tool for PMB exposure prediction, particularly with limited sampling data.