Comparison of population pharmacokinetic modeling and machine learning approaches for predicting voriconazole trough concentrations in critically ill patients.

Abstract

Background: Despite the widespread use of voriconazole in antifungal treatment, its high pharmacokinetic and pharmacodynamic variability may lead to suboptimal efficacy, especially in intensive care unit (ICU) patients. Machine learning (ML), an artificial intelligence modeling approach, is increasingly being applied to personalized medicine. The effectiveness of ML models for predicting voriconazole blood concentrations in ICU patients, compared to traditional population pharmacokinetics (popPK) models, has been uncertain until now. This study aims to identify the most effective modeling strategy for voriconazole.

Methods: We developed six ML models using 244 concentrations from 62 patients in our previous popPK dataset. Another additional dataset, consisting of 282 trough concentrations from 177 patients, was used to externally evaluate both ML models and five other published popPK models, utilizing prediction-based diagnostics, simulation-based diagnostics, and Bayesian forecasting.

Results: The XGBoost model exhibited superior predictive performance among the six ML models, achieving an R² of 0.73. Its performance metrics (RMSE%: 127.21 %, median absolute prediction error: 29.65 %, median prediction error: 9.82 %, F20: 34.04 %, F30: 50.71 %) outperformed those of the best popPK model (RMSE%: 152.41 %, median absolute prediction error: 44.75 %, median prediction error: -0.99 %, F20: 23.40 %, F30: 36.88 %), suggesting greater accuracy and precision in predicting pharmacokinetics.

Conclusions: Both ML and popPK models can be utilized for individualized voriconazole therapy. Our comparative study provides insights into the most effective methods for modeling and predicting voriconazole concentrations.