Category: Clinical Pharmacology
Purpose: 1) To compare the performance of first-order conditional estimation (FOCE) and two different expectation-maximization (EM) methods, Stochastic Approximation EM (SAEM) and Monte Carlo Importance Sampling Parametric EM (IMP), and 2) to investigate the efficiency of parallel computing in the development of a population pharmacokinetic (PopPK) model of oxfendazole and metabolites following single ascending doses of oxfendazole in healthy adults.
Methods: Data were obtained from the first-in-human single ascending doses study of oxfendazole in healthy adult volunteers. Model development, including all algorithm evaluated, was performed In NONMEM 7.4.2. Non-compartmental analysis (NCA) was performed using Phoenix WinNonlin 8.0. Oxfendazole PK was characterized using a one-compartment structural model with nonlinear absorption process. Two other compartments, one for the major metabolites, oxfendazole sulfone, and one for the minor metabolite, fenbendazole, were included. Both pre-systemic and systemic metabolisms were necessary to sufficiently characterized oxfendazole sulfone PK; meanwhile, fenbendazole disposition was well captured by first-order systemic metabolism. Absorption rate constant, metabolic rate constants and elimination rate constants were first-order (Figure 1). Exponential inter-individual variability along with proportional and additive residual variability were applied. To compare the performance among the three estimation methods, following items were evaluated: 1) the average difference between individual oxfendazole exposure (i.e. AUC and Cmax) and final parameter estimates obtained from model fitting and those obtained from NCA, 2) precision of final parameter estimates, 3) model stability and 4) computational time. For each estimation method, model development was repeated with variable number of CPUs to assess the extent of increase in model development efficiency given the increase number of CPUs in parallel computing platform.
Results: Even though there was not much difference in final estimates of population parameters (Figure 2), results from FOCE was unstable and unreliable with eigenvalues ratio being greater than 1000 and extreme precisions (i.e. %CV ranging from almost 0 to more than 1000). IMP and SAEM yielded almost the same final parameter and precision estimates. IMP was twice as fast as SAEM, while FOCE took the longest computational time (Figure 3). Computational speed increased approximately proportionally with the increase in number of CPUs. However, for FOCE, the difference between single CPU platform and parallel computing platform was significantly more than proportional.
Conclusion: Given a rich data set, yet complicated structural model, IMP method was the most efficient in terms of time and precision for PopPK model development.