Machine Learning-Based Model Selection and Averaging Outperform Single-Model Approaches for a Priori Vancomycin...

Area of Science:

• Pharmacometrics
• Machine Learning in Healthcare
• Drug Dosing Optimization

Background:

• Selecting optimal population pharmacokinetic (PK) models for individual patients in model-informed precision dosing (MIPD) is challenging, especially without therapeutic drug monitoring (TDM) data.
• Vancomycin dosing requires careful consideration due to its narrow therapeutic index and potential for toxicity.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model to guide the selection of appropriate PK models for a priori MIPD of vancomycin.
• To improve the accuracy and reliability of vancomycin dosing predictions in the absence of TDM samples.

Main Methods:

• A retrospective analysis of 343,636 vancomycin TDM records from adult patients across 156 healthcare centers.
• Development of an ML multi-label classification model (XGBoost) to predict the performance of six existing PK models.
• PK model predictions were labeled based on their proximity (80%-125%) to observed TDM values.

Main Results:

• The ML-guided selection of the highest-ranked PK model and ML-based model averaging significantly outperformed individual PK models, BMI-based selection, and naive averaging.
• ML approaches demonstrated improved population-level accuracy, a higher proportion of predictions within the target range, and no systematic bias.
• Predictive performance decreased with lower ML-assigned model rankings, indicating the utility of the ML-driven ranking system.

Conclusions:

• ML-driven PK model selection and averaging provide a robust strategy for a priori MIPD of vancomycin.
• These ML approaches can enhance early vancomycin dosing decisions by guiding the selection of suitable models and avoiding suboptimal ones.
• The findings suggest that ML can optimize precision dosing in scenarios with limited or no TDM data.