Assessing Covariate Clinical Relevance in High‐Dimensional <scp>PK</scp> Analysis: A Comparison of <scp>SCM</scp> +, <scp>FFEM</scp> , and <scp>FREM</scp> Approaches

This work aimed to assess the correctness of covariate clinical relevance (CCR) assessment using SCM+, FFEM, and FREM within a high-dimensional covariate framework with varying effect sizes and correlations. A clinical trial simulation inspired by the dupilumab case study was conducted (200 datasets of 300 patients each), using a 2-compartment PK model with 12 covariates having small, medium, or high effect size. Covariate analysis was based on a 70 covariate-parameter relationships predefined set including 12 continuous and 7 binary covariates sampled from the NHANES database, spanning high to low correlations. The simulated reference model (RM) was fitted for comparison. Parameter estimation was performed in NONMEM/PsN using FOCEi (SCM+, FFEM) or IMPMAP (FREM), with SE derived from the S matrix. CCR assessment followed a forest plot-inspired approach: 90% confidence intervals with a [0.8-1.25] reference area for clinical relevance; 5% type I error for statistical significance. Parameter estimates and SE were always obtained, allowing full CCR evaluation. For covariates with simulated effects, all methods were consistent with RM. SCM+ missed up to 9% of small-effect covariates, whereas FFEM/FREM more often indicated insufficient information to conclude across all effect sizes. For covariates without any effect, SCM+ mostly did not select them, while FFEM/FREM was more informative by classifying them as non-relevant or with insufficient information. As non-selection may reflect a lack of power rather than no effect, robust CCR assessment should begin with FFEM/FREM for a comprehensive exploration, followed by SCM+ to build a parsimonious predictive model.