Morgane Philipp1,2, Sylvie Retout2,3, Simon Buatois3, France Mentré1
1Université Paris Cité, INSERM, IAME, UMR 1137, 2Institut Roche, 3Roche Pharma Research and Early Development, Pharmaceutical Sciences, Roche Innovation Center Basel
Objectives Covariate analysis is key in population PK modeling to identify and quantify variability between individuals and guide dose adjustments, reducing under- or overexposure risks. Covariate model building methods fall into 2 families: covariate selection methods, e.g. SCM [1] and SCM+ [2], studying the effect of selected covariates; and full modeling methods, e.g. FFEM [3] and FREM [4,5], studying all covariate effects of the predefined set without selection. Once the covariate model is developed, covariate impacts on drug exposure can be visualized using forest plots [6]. It represents covariate effect ratios (CER) and their 90% confidence intervals (CI90) relative to a typical individual. Covariate clinical relevance (CCR) can be inferred from the position of CER and CI relative to the reference area [0.8, 1.25] and reference line at 1 [7]. Accurate CCR assessment is essential for a proper evaluation of dose adjustment needs. Published studies comparing covariate modelling methods mainly focus on covariate selection performance and estimation accuracy, without evaluating the CCR [8]. In addition, no study comparing SCM and FFEM has been conducted yet. The objectives of this work were: 1. Evaluate and compare the adequacy of CCR decisions in a population PK analysis using FFEM, SCM and SCM+ with 7 covariates (5 continuous, 2 categorical) and 14 predefined relationships (project 1) 2. Extend the comparison to a more complex framework, incorporating a high-dimensional covariate set with 19 covariates (12 continuous, 7 categorical) and 70 predefined relationships, and including FREM for a comprehensive evaluation (project 2) Methods SCM (project 1), SCM+ (both projects), FFEM (both projects) and FREM (project 2) algorithms implemented in PsN were compared for CCR assessment using clinical trial simulation. 200 datasets were simulated for project 1 and 50 for project 2. Parameter estimation was done using NONMEM with FOCEi for SCM/SCM+, FFEM and IMPMAP (NITER=3000) for FREM to overcome estimation issues [5]. SE were derived from the variance-covariance matrix computed as R-1SR-1. FREM covariate effects were calculated univariately (uni) or multivariately (multi) [4] and their SE were derived from 1000 draws. By design, FREM tested all covariate/parameter combinations, with a total of 95 relationships for project 2. Cholesky decomposition can be used to force the relationships not included in the predefined set to 0 [10]. CCR evaluation was done using the proposed classification [9]: relevant (R), non-relevant significant & non-significant (NRS/NRNS), insufficient information significant & non-significant (IIS/IINS) depending on whether CER and CI were outside, within, or straddling the reference area without/with crossing the reference line. The simulated model, i.e. reference model (RM), was fitted to each dataset as a reference for CCR assessment. For project 2, clinical significance was evaluated using IC95 to ensure a common 5% threshold. OFV, BICc, and computation time were also compared to evaluate performance. For project 1, a clinical trial inspired from a real case study of 383 hemophilia A patients treated with emicizumab [11], combining rich and sparse sampling designs was simulated. Covariate distributions were based on real data and a 1-compartment model with first-order absorption and linear elimination was used, simplified from Retout et al [10]. 2 scenarios were investigated: (i) a “base model” including a body weight (BW) effect on CL/F and V/F, (ii) a “covariate model” including four additional relationships. Upon the simulated relationships, BW had a moderate effect (CER=1.4) while the others had small ones (0.8<CER<1.2). For project 2, the NHANES database was used for covariate simulations. A 300-patient study with a rich sampling design was simulated using a 2-compartment model with first-order absorption and linear elimination [12]. The covariate model included 12 simulated relationships with varying effect sizes: small (CER=1.25, n=4 relationships), moderate (CER=1.4, n=5) and strong (CER=1.6, n=3). Results In project 1, for relationships with a simulated effect, FFEM, SCM and SCM+ yielded consistent results in line with those of the RM. BW effect was found R in almost 100% of cases. However, significant covariates were missed in up to 15% of cases with SCM/SCM+. For relationships without a simulated effect, FFEM mostly identified them NR or II, whereas SCM/SCM+ mainly did not select them (with the few additional ones retained being NR or II). In project 2, strong effect relationships were always found R (100%) with RM and SCM+, while FFEM reported them R (84-100%), IIS (0-12%) or IINS (0-4%). Moderate effect relationships were mostly found R (66-100%) or IIS (0-34%) with the RM and SCM+. FFEM gave similar results but showed discrepancies for 1 relationship, detected as R (46%), IIS (48%) or IINS (6%), instead of 66%, 34% and 0%, respectively for RM. Low effect relationships were mostly found IIS (56%-88%), R (10-44%), and rarely NRS (0-6%) or NRNS (0-2%) with the RM. SCM+ gave similar results except in up to 9% of the cases where these relationships were not selected. FFEM showed a high proportion of IINS (0-66%) and NRNS (0-12%). FREM uni, gave results roughly like those of FFEM but with an higher IINS proportion while FREM multi found them mainly IINS, sometimes R, rarely IIS or NRNS. For relationships without a simulated effect, SCM+ mainly did not select them; additional ones were retained in 44% of the cases and either found NR or II. FFEM identified them as NR or II with a higher proportion of II for categorical covariates (6-100%) compared to continuous ones (2-82%). FREM uni also found them mainly II or NR, occasionally R while with FREM multi they were mainly IINS and rarely NRNS. FFEM outperformed SCM+ in terms of OFV (medians=48512 vs 48573) but not in terms of BICc (medians=48792 vs 49261) due to more parsimonious models. Finally, FFEM was 10 times faster than SCM+ (median runtimes=3 vs 33.7h). FREM had a median runtime of 54 hours. Conclusion FFEM, SCM, and SCM+ provided satisfactory CCR assessment for a standard-dimensional covariate simulation framework, but challenges emerged for full covariate modeling methods in high-dimensional settings. Larger SEs were obtained for FFEM and FREM due to a higher number of parameters to be estimated. FREM results remain partially interpretable as FREM multi tested 25 more relationships than FFEM, requiring Cholesky decomposition for fair comparison. SCM/SCM+ miss significant covariates in up to 15% of cases in each simulation, but only for low effect relationships that were not R. Assuming that non-selected covariates have no effect remains an assumption unnecessary with full modeling methods. To ensure a robust covariate analysis, a full modeling method and a covariate selection method should both be reported to get the CCR evaluation of all relationships and a parsimonious model suitable for prediction.
[1] Ahamadi M, Largajolli A, Diderichsen PM, et al (2019) Operating characteristics of stepwise covariate selection in pharmacometric modeling. J Pharmacokinet Pharmacodyn 46:273–285. https://doi.org/10.1007/s10928-019-09635-6 [2] Svensson RJ, Jonsson EN (2022) Efficient and relevant stepwise covariate model building for pharmacometrics. CPT Pharmacomet Syst Pharmacol 11:1210–1222. https://doi.org/10.1002/psp4.12838 [3] Xu XS, Yuan M, Zhu H, et al (2018) Full covariate modelling approach in population pharmacokinetics: understanding the underlying hypothesis tests and implications of multiplicity. Br J Clin Pharmacol 84:1525–1534. https://doi.org/10.1111/bcp.13577 [4] Yngman G, Bjugård Nyberg H, Nyberg J, et al (2022) An introduction to the full random effects model. CPT Pharmacomet Syst Pharmacol 11:149–160. https://doi.org/10.1002/psp4.12741 [5] Jonsson EN, Nyberg J (2024) Full random effects models (FREM): A practical usage guide. CPT Pharmacometrics Syst Pharmacol 13: 1297-1308. https://doi.org/10.1002/psp4.13190 [6] Menon-Andersen D, Yu B, Madabushi R, et al (2011) Essential Pharmacokinetic Information for Drug Dosage Decisions: A Concise Visual Presentation in the Drug Label. Clin Pharmacol Ther 90:471–474. https://doi.org/10.1038/clpt.2011.149 [7] Cuzick J (2005) Forest plots and the interpretation of subgroups. The Lancet 365:1308. https://doi.org/10.1016/S0140-6736(05)61026-4 [8] Karlsen M, Khier S, Fabre D, et al (2025) Covariate Model Selection Approaches for Population Pharmacokinetics: A Systematic Review of Existing Methods, From SCM to AI. CPT Pharmacometrics Syst Pharmacol. https://doi.org/10.1002/psp4.13306 [9] Philipp M, Buatois S, Retout S, et al (2024) Impact of covariate model building methods on their clinical relevance evaluation in population pharmacokinetic analyses: comparison of the full model, stepwise covariate model (SCM) and SCM+ approaches. J Pharmacokinet Pharmacodyn https://doi.org/10.1007/s10928-024-09911-0 [10] Philipp M, Buatois S, Retout S, et al (2024) Extension of the covariate clinical relevance determination in population pharmacokinetic analyses to the Least Absolute Shrinkage and Selection Operator (LASSO) and to Genetic Algorithm (GA) approaches. PAGE 32, Roma, Italy www.page-meeting.org/?abstract=11031 [11] Retout S, Schmitt C, Petry C, et al (2020) Population Pharmacokinetic Analysis and Exploratory Exposure–Bleeding Rate Relationship of Emicizumab in Adult and Pediatric Persons with Hemophilia A. Clin Pharmacokinet 59:1611–1625. https://doi.org/10.1007/s40262-020-00904-z [12] Kovalenko P, Davis JD, Li M et al (2020) Base and Covariate Population Pharmacokinetic Analyses of Dupilumab Using Phase 3 Data. Clinical Pharmacology in Drug Development, 9: 756-767. https://doi.org/10.1002/cpdd.780
Reference: PAGE 33 (2025) Abstr 11629 [www.page-meeting.org/?abstract=11629]
Poster: Methodology - New Modelling Approaches