Peng Man1, Yevgen Ryeznik1, Xiaomei Chen1, Rikard Nordgren1, Stella Belin1, Chee M. Ng2, Robert J. Bauer3, Mats O. Karlsson1, Andrew C. Hooker1
1Department of Pharmacy, Uppsala University, 2NewGround Pharmaceutical Consulting LLC, 3Pharmacometrics, R&D, ICON Clinical Research, LLC
Introduction: Stepwise covariate modeling (SCM) offers an automated, data-driven approach for covariate model selection via forward inclusion and backward elimination (1). However, its computational efficiency is limited by the need to evaluate numerous covariate effects through nonlinear mixed effects models. To address this, Stochastic Approximation for Model Building Algorithm (SAMBA) was proposed (2). SAMBA uses linear covariate models with Monte Carlo samples of random effects to screen covariate effects, reducing nonlinear model evaluations. Furthermore, by leveraging sampled random effects from SAEM-derived ETA distributions, SAMBA is believed to be less prone to the impact of high shrinkage compared to empirical Bayes estimates (EBEs) (3). In this study, we applied EBE-based automated linear covariate screening (LCS) to SCM and developed two methods: SCM using stepwise LCS (SCM-SLCS) and SCM using non-stepwise LCS (SCM-LCS). Objectives: -evaluate SCM-SLCS and SCM-LCS along with standard SCM and SAMBA under varying ETA shrinkage levels; -explore approaches to account for EBE uncertainty in LCS, such as weighted least squares (WLS) using ETCs or linear mixed effects models (LME) using posterior sampling from the FOCE-derived ETA distribution. Methods: Datasets were simulated using a one-compartment PK model with 1st-order absorption and linear elimination. 6 scenarios were created by varying sampling designs (rich: 5 samples, N5; sparse: 2–3 samples, N2 and N3) and adjusting the IIV-to-RUV ratio (4). Two parameter sets were used: Set 1 (?CL=?VC=0.5, proportional s=0.3) and Set 2 (?CL=?VC=0.2, proportional s=0.4), with shared parameters (MAT=0.1 h, CL=0.5 L/h, VC=0.2 L, ?MAT=0.1, additive s=0.15). Covariates were sampled from the 2017–2020 NHANES database. A true covariate effect of weight on VC was implemented with a power model (VCWT=1.0 for S1, 0.5 for S2). The impact of reduced effect size was also assessed (VCWT=0.5 for S1, 0.2 for S2). For each scenario, 500 replicate datasets with 500 individuals each were generated. The effectiveness of covariate modeling methods in identifying the correct parameter-covariate relationships was evaluated using power, type I error rate, true positive rate, true negative rate, and inclusion frequency. Computational efficiency was assessed via runtime. Unless otherwise stated, only forward inclusion was carried out. All methods were implemented in Pharmpy (5) and NONMEM. Results: ETA shrinkage in this simulation setup ranged from 7.34% to 88.8%. When tested with a single covariate effect, SCM, SCM-SLCS, and SCM-LCS achieved high power (>90%) and controlled type I error rate (~4%) across all scenarios. SAMBA’s power was high in low shrinkage scenarios but declined as shrinkage increased, dropping to 88.4% in S2N3 and 34.0% in S2N2. Notably, SAMBA maintained a low type I error rate (0.6%) across all scenarios. Reducing the covariate effect magnitude decreased power across all methods. SCM’s power declined from 99.0% to 97.2% in S1N5 (lowest shrinkage) and from 99.8% to 50.6% in S2N2 (highest shrinkage). SCM-SLCS and SCM-LCS showed similar trends, with power ranging from 98.2% in S1N5 to 48.4% in S2N2. SAMBA was the most affected, with power decreasing from 100% to 87.6% in S1N5 and from 34.0% to 0% in S2N2. While searching multiple covariate effects, SCM-LCS maintained high sensitivity (96%) and improved specificity by 6–11% compared to SCM. SAMBA exhibited the highest specificity, but its sensitivity was more noticeably reduced in high shrinkage scenarios. Average runtimes (±SD) were 12.94±5.02 (SCM), 17.48±5.45 (SAMBA), 4.12±0.79 (SCM-SLCS), and 5.37±1.75 (SCM-LCS) minutes. SCM-LCS supported introducing individual uncertainty using WLS with ETCs or using LME with posterior samples, but these approaches showed no improvement over ordinary least squares in tested scenarios. Conclusions: We developed SCM-SLCS and SCM-LCS, integrating EBE-based LCS into standard SCM. Under a unified implementation and testing framework (Pharmpy), SCM-LCS demonstrated comparable sensitivity to SCM, improved specificity, and significantly reduced runtime, making it a promising method for efficient covariate selection. SAMBA, while demonstrating high specificity, exhibited reduced sensitivity in response to high shrinkage and low effect magnitude. Additionally, SCM-LCS provides a framework to integrate individual uncertainty using WLS or LME, potentially allowing for even more robust covariate selection.
1. Jonsson EN, Karlsson MO. Automated covariate model building within NONMEM. Pharm Res. 1998 Sep;15(9):1463–8. 2. Prague M, Lavielle M. SAMBA: A novel method for fast automatic model building in nonlinear mixed-effects models. CPT: Pharmacometrics & Systems Pharmacology. 2022;11(2):161–72. 3. Lavielle M, Ribba B. Enhanced Method for Diagnosing Pharmacometric Models: Random Sampling from Conditional Distributions. Pharm Res. 2016 Dec;33(12):2979–88. 4. Combes FP, Retout S, Frey N, Mentré F. Powers of the Likelihood Ratio Test and the Correlation Test Using Empirical Bayes Estimates for Various Shrinkages in Population Pharmacokinetics. CPT Pharmacometrics Syst Pharmacol. 2014 Apr;3(4):e109. 5. Chen X, Nordgren R, Belin S, Hamdan A, Wang S, Yang T, et al. A fully automatic tool for development of population pharmacokinetic models. CPT: Pharmacometrics & Systems Pharmacology. 2024;13(10):1784–97.
Reference: PAGE 33 (2025) Abstr 11560 [www.page-meeting.org/?abstract=11560]
Poster: Methodology - New Modelling Approaches