Utilization of Machine Learning Approaches for Drug Clearance Prediction and Population Pharmacokinetic Covariate Analysis

Abstract

Population pharmacokinetic (popPK) analysis is routinely used to evaluate drug clearance and covariate effects on pharmacokinetic parameters to support dosing recommendations. While stepwise methods are traditionally employed for covariate identification, artificial intelligence (AI) and machine learning (ML) approaches offer promising alternatives for enhancing these analyses. This proof-of-concept study illustrates the application of AI/ML and explainable artificial intelligence (XAI) techniques for drug clearance prediction and covariate analysis using two distinct datasets for the drugs methotrexate and remifentanil. For the larger methotrexate dataset, we utilized multiple ML models including convolutional neural networks, logistic regression, and gradient boosting and highlighted exceptional performance (R² for accuracy > 0.96) in clearance prediction. XAI via SHapley Additive exPlanations (SHAP) analysis is utilized to identify vital covariates impacting clearance. Here, XAI techniques are utilized to explore how different AI/ML approaches might impact the interpretation of relationships among covariates. By examining these methods, we seek to better understand their respective strengths, limitations, and potential to provide insights for popPK analysis. The second example used a smaller dataset for the drug remifentanil and included pediatric to adult populations. Here, the performance of the ML models was more modest (maximum R² of 0.75), highlighting the dependence of ML techniques on adequate sample sizes. SHAP analysis confirmed age and weight as critical covariates for the clearance of remifentanil. Our findings demonstrate that AI/ML approaches can provide accurate clearance predictions and identify potentially overlooked covariates in an unbiased, hypothesis-free manner. However, this study also emphasizes important limitations, including the requirement for sufficiently large datasets and the drug-specific nature of trained models. These proof-of-concept examples illustrate how AI/ML methods can complement traditional pharmacokinetic analyses, offering additional insights while maintaining scientific rigor.