Evaluation of the Lindstrom-Bates FOCE Algorithm with Simulated Pharmacokinetic Datasets
J. Chittenden, R. H. Leary, B. Matzuka, M. Dunlavey
Pharsight Corporation, Cary, NC, USA
Objective: Accuracy, run time, and robustness are primary concerns when selecting a method for estimation of a population pharmacokinetic model. Historically, the most popular methods have been the First Order Conditional Estimation (FOCE) methods provided by NONMEM®, which generally provide acceptable accuracy in parameter estimates and likelihood evaluation. An alternative FOCE algorithm was proposed by Lindstrom and Bates [1] that allows a significant simplification to the optimization of the likelihood and should result in faster runtimes. To our knowledge, an extensive investigation of the accuracy, speed, and robustness of the Lindstrom-Bates FOCE (FOCE-LB) algorithm with respect to pharmacokinetic data has not been conducted. This work compares the FOCE-LB algorithm implemented in Phoenix® NLME 1.1 with NONMEM VII FOCE (FOCE-ELS) results.Methods: For this evaluation we use a large set of test data and models that were previously generated by Laveille et al [2] for an evaluation of the SAEM algorithm implemented in MONOLIX. The models were transcribed to Pharsight Modeling Language (PML) for use with Phoenix NLME and both Phoenix NLME and NONMEM model sets were automated in the same environment with equivalent settings and initial estimates. We compare run time of the main algorithm (excludes covariance estimate), ELS log-likelihood to ascertain quality of convergence, convergence code or message, and parameter estimates. We compare cases where FOCE-ELS and FOCE-LB both converge and highlight cases where convergence is different.
Results:
- The FOCE-LB algorithm obtains comparable accuracy in 94% of the cases, compared to 73% for FOCE-ELS.
- The FOCE-LB algorithm was faster in over 94% of the cases where it converged. On average it was 4 times faster than FOCE-ELS and it was 13 times faster over all the mutually converged problems.
- The FOCE-LB algorithm converged in over 95% of the cases in which FOCE-ELS also converged. Of these, 21% converged to a different and significantly worse result and 15% converged to a different and significantly better result than FOCE-ELS.
- The FOCE-ELS algorithm converged in 77% of the cases in which FOCE-LB also converged. Of these, 25% converged to a different and significantly worse result and 21% converged to a different and significantly better result than FOCE-ELS.
Conclusions: The FOCE-LB algorithm as implemented Phoenix NLME is a fast, accurate, and reliable method for estimating population pharmacokinetic models in the 150 Monolix test cases investigated here.
References:
[1] Nonlinear Mixed Effects Models for Repeated Measures Data, M. Lindstrom, D. Bates, Biometrics , Vol. 46, No. 3 (Sep., 1990), pp. 673-687
[2] C. Laveille, M. Lavielle, K. Chatel, P. Jacqmin; Evaluation of the PK and PK-PD libraries of MONOLIX: A comparison with NONMEM; PAGE 2008
