Use Machine Learning for Covariate Selection in Population Model

Covariate selection is a critical step in developing robust population models, influencing their predictive accuracy and generalizability. This study evaluates three covariate selection methods: (a) stepwise covariate selection with expert pharmacometrician (PMx) input, (b) the Perl-Speaks-NONMEM Stepwise Covariate Model (PsN-SCM) tool, and (c) a machine learning (ML) approach leveraging SHAP values to infer optimal covariate functional forms. Among these methods, the ML-based approach demonstrated superior computational efficiency and rapid covariate selection. However, the PMx-selected model exhibited the best performance in both internal and external validation, highlighting the continued importance of expert domain knowledge in model development. In this talk, we will delve into the details of performing covariate selection in ML models and evaluating ML model performance. We will also provide Python code examples to demonstrate the implementation of ML-based covariate selection techniques. In the end, we will briefly discuss the advantages of ML methods in covariate selection and the key considerations pharmacometricians should keep in mind when adopting this novel approach. [1] Janssen, Alexander, et al. "Application of SHAP values for inferring the optimal functional form of covariates in pharmacokinetic modeling." CPT: Pharmacometrics & Systems Pharmacology 11.8 (2022): 1100-1110.