Comparing the performance of first-order conditional estimation (FOCE) and different expectation–maximization (EM) methods in NONMEM: real data experience with complex nonlinear parent-metabolite pharmacokinetic model

First-order conditional estimation (FOCE) has been the most frequently used estimation method in NONMEM, a leading program for population pharmacokinetic/pharmacodynamic modeling. However, with growing data complexity, the performance of FOCE is challenged by long run time, convergence problem and model instability. In NONMEM 7, expectation–maximization (EM) estimation methods and FOCE with FAST option (FOCE FAST) were introduced. In this study, we compared the performance of FOCE, FOCE FAST, and two EM methods, namely importance sampling (IMP) and stochastic approximation expectation–maximization (SAEM), utilizing the rich pharmacokinetic data of oxfendazole and its two metabolites obtained from the first-in-human single ascending dose study in healthy adults. All methods yielded similar parameter estimates, but great differences were observed in parameter precision and modeling time. For simpler models (i.e., models of oxfendazole and/or oxfendazole sulfone), FOCE and FOCE FAST were more efficient than EM methods with shorter run time and comparable parameter precision. FOCE FAST was about two times faster than FOCE but it was prone to premature termination. For the most complex model (i.e., model of all three analytes, one of which having high level of data below quantification limit), FOCE failed to reliably assess parameter precision, while parameter precision obtained by IMP and SAEM was similar with SAEM being the faster method. IMP was more sensitive to model misspecification; without pre-systemic metabolism, IMP analysis failed to converge. With parallel computing introduced in NONMEM 7.2, modeling speed increased less than proportionally with the increase in the number of CPUs from 1 to 16.