Beyond Traditional Covariates: An Interpretable Machine Learning Workflow for Improved Hybrid Pharmacometric Modeling

Area of Science:

• Pharmacometrics
• Machine Learning
• Clinical Pharmacology

Background:

• Traditional population pharmacokinetic (PPK) models struggle with generalizability, especially in critically ill patients.
• Accurate a priori pharmacokinetic predictions are crucial for effective and safe drug dosing.

Purpose of the Study:

• To develop and validate a hybrid machine learning-PPK framework for improved pharmacokinetic predictions.
• To enhance individualized dosing strategies by integrating real-world clinical data.

Main Methods:

• A hybrid framework combined two-compartment PPK models with machine learning (XGBoost, symbolic regression).
• Maximum a posteriori Bayesian estimation adjusted population parameters using MIMIC-IV data.
• Machine learning models predicted pharmacokinetic parameters (clearance, volume of distribution) for a priori vancomycin trough concentration prediction.

Main Results:

• The hybrid models significantly improved prediction accuracy compared to traditional PPK covariate models.
• Vancomycin trough concentration prediction error was reduced by up to approximately 20%.
• XGBoost showed high predictive performance, while symbolic regression offered interpretable insights into predictor-parameter relationships.

Conclusions:

• Combining machine learning with PPK modeling refines pharmacokinetic parameter estimation for precise, individualized dosing.
• This adaptable framework can identify non-traditional predictors and enhance pharmacometric model performance across various drugs and populations.