A Multi-Method Comparison of Machine Learning in Predicting Pharmacokinetic Parameters: A Simulation Study

Abstract

Introduction: One important aim of population pharmacokinetics (PK) and pharmacodynamics (PD) is the identification and quantification of the relationships between the parameter and covariates to improve the predictive performance of the population PK/PD modeling. Several new mathematical methods have been developed in pharmacokinetics in recent years which indicated that the machine learning-based methods are an appealing tool for analyzing PK/PD data. Methods: This simulation-base study aims to determine whether machine learning methods, including support vector regression (SVR) and Random forest (RF) which are specifically designed for the prediction of blood serum concentration or clearance, could be an effective replacement for the Lasso covariate selection method in nonlinear mixed effect models. Accordingly, the predictive performance of penalized regression Lasso, SVR, and RF regression was compared to detect the associations between clearance and model covariates. PK data was simulated from a one-compartment model with oral administration. Covariates were created by sampling from a multivariate standard normal distribution with different levels of correlation. The true covariates influenced only clearance at different magnitudes. Lasso, RF, and SVR were compared in terms of mean absolute prediction error (MAE). Results: The results show that SVR performed the best in small data sets, even in those in which a high correlation existed between covariates. This makes SVR a promising method for covariate selection in nonlinear mixed-effect models. Conclusion: The Lasso method offered a higher MAE, making it less promising than RF and SVR, especially when dealing with a high correlation between covariates and a low number of individuals.