Managing Director NewGround Pharmaceutical Consulting LLC Foster City, California, United States
Objective: Development of optimal covariate model for population data analysis is time-consuming, labor-intensive, and nontrivial process. Among many different approaches used to develop covariate models, stepwise selection method (SCM) with forward selection/backward elimination (FS/BE) is most commonly used in population pharmacokinetic/pharmacodynamic (PPK/PD) analysis. Although this standard method can reduce number of models to be tested, it is still time-consuming and labor-intensive when testing many different models and complex models with long run times. Hence, aim of this study is to develop four fast and efficient SCM methods using different covariate-parameter relationship (C-P) screening approaches and compared performance of these models to standard SCM (S-SCM) in covariate model building of PPK analysis.
Method: Four fast SCM with FS/BE methods using different C-P relationships screening approaches were developed 1) FF-SCM: All C-P relationships that meet screening criteria were included in separate model run for each FS process of SCM; 2) FS-SCM: Included all C-P relationships that meet screening criteria in a single model run for each FS process of SCM; 3) FFS-SCM: Included only most significant C-P relationship in model run for each FS process of SCM; and 4) C-SCM: Used Monolix COSSAC-like C-P screening approach to select C-P relationships in model run for each iteration of SCM. One hundred simulated datasets using one-compartment linear PK model with intensive/sparse PK sampling design were used to compare performance of four fast SCM methods with S-SCM methods. Nominal models were used as references and sensitivity/specificity were computed to evaluate ability of these models to accurately identify C-P relationships. Model run times were compared. Selection of C-P relationships in screening process was based on pre-determined significance levels (0.1 and 0.05). Two pre-determined significance levels were used to select C-P for FS (PFS) and BE (PBE) processes of SCM method: PFS=0.05 with PBE=0.01 and PFS=0.01 with PBE=0.01.
Results: For 100 simulated datasets with intensive PK sampling, all fast SCM methods achieved shorter run times compared to S-SCM. FFS-SCM had shorter average run times (2.92 min) than FF-SCM (6.03 min) and FS-SCM (9.39 min) and achieved similar accuracy (sensitivity and specificity >90%) compared to S-SCM (15.5 min). C-SCM had shortest average run times (1.88 min) but noticeable different results in its ability to include right/wrong covariates in model compared to S-CSM. Comparable results were observed for simulated datasets with sparse PK sampling.
Conclusions: Four fast SCM methods with noticeable shorter model run times than S-SCM were developed, and FFS-SCM had shorter run times than FF-SCM and FS-SCM while achieved similar accuracy compared to S-SCM. FFS-SCM is useful to develop covariate model in population data analysis for improving efficiency of model-based drug development.
Citations:
References: [1]. Ayral G, Abdallah JFS, Magnard C, Chauvin J et al. CPT Pharmacometrics Syst Pharmacol. 2021;10(4):318-329 [2]. Zou Y, Tang F, Ng CM.AAPS J 2021;23(2):37
