NONMEM and NPEM2 population modeling: a comparison using tobramycin data in neonates.

Nonlinear mixed effects modeling (NONMEM) and nonparametric expectation maximization (NPEM2) have both been used in population modeling of tobramycin. We compared both methods for differences in population pharmacokinetic parameters in relation to error models used. Predictive performance was compared between models. A group of 470 neonates who had received tobramycin according to a gestational age (GA)-dependent dosing interval was analyzed according to a one-compartment model with NONMEM and NPEM2. Additional models were constructed where the assay error pattern in NPEM2 mimicked NONMEM residual error and vice versa. Individual pharmacokinetic parameter estimates were compared. Predictive performance was evaluated in a separate group of 61 patients. Population estimates and variation coefficients (CV) for optimal models were NONMEM K(el) 0.071 h(-1) (27%), V(d) 0.59 L/kg (9%); NPEM2 K(el) 0.079 h(-1) (42%), V(d) 0.65 L/kg (48%). Forcing NONMEM to use the NPEM2 error pattern as residual error or vice versa resulted in smaller differences in CVs of the estimates. NONMEM gave less bias (P < 0.05) than NPEM2 and comparable precision with this approach. In conclusion NONMEM and NPEM2 are dissimilar in population estimates. Differences in ranges of pharmacokinetic parameter estimates between NONMEM and NPEM2 are largely determined by the method of incorporating error patterns in both programs.