Shap-Cov: An Explainable Machine Learning Based Workflow for Rapid Covariate Identification in Population Modeling

Covariate identification in population pharmacokinetic/pharmacodynamic (popPK/PD) modeling is a key component in model development that is often prone to bias, time-consuming, and even intractable when too many covariates or complicated models are being considered. Early work leveraging machine learning (ML) for covariate screening has shown promising results over traditional methods. In this work, we expand this effort by integrating explainable machine learning facilitated by Shapley Additive Explanations (SHAP) analysis and covariate uncertainty quantification as well as a formal framework for establishing statistical significance of covariate relationships. Finally, we have packaged the proposed methodology into a flexible set of functions (shap-cov) to support popPK/PD modeling covariate identification.