Development of the population PK model.

The population PK model was built using nonlinear mixed-effects modeling via NONMEM software v7.3 (Icon Development Solutions, Dublin, Ireland) in conjunction with Perl-speaks-NONMEM (PsN) 3.5.3 (25). Compartmental PK models were coded using ADVAN1 TRANS2 and ADVAN3 TRANS4 subroutines in NONMEM. Data exploration, manipulation, and graphics were handled using Xpose 4.3.5 embedded in R 3.1.0 (http://cran.r-project.org/; R is an open-source, S-based statistical software) (26, 27). The first-order conditional estimation with interaction method (FOCE-I) was used to estimate PK parameters. The variability of the parameter was estimated using the covariance step. Visual inspections of the amikacin blood concentration-versus-time profiles and the objective function value (OFV), calculated using likelihood ratio tests, were used to determine the base model. A heteroscedastic model (proportional) was selected to describe the interindividual variability in amikacin PK parameters, and the residual variability was modeled as a homoscedastic error model (additive).

Following model development, covariates were evaluated in a stepwise forward selection, and significant covariates were combined in a full model to characterize amikacin PK. This was followed by backward elimination, and significant covariates were retained in the final model. The continuous covariates tested were age, total body weight, adjusted and ideal body weight, body mass index (BMI), creatinine, creatinine clearance (estimated using the Cockcroft-Gault equation with total body weight [28], as well as CKD-EPI [29] and MDRD [30] estimations), urea, blood urea nitrogen, total proteins, albumin, total bilirubin, serum electrolytes (sodium, potassium, chloride, calcium, phosphorus, and magnesium), and acute physiology and chronic health evaluation II (APACHE II) scores. The categorical covariates evaluated were sex, mechanical ventilation, diabetes mellitus, arterial hypertension diagnosis, and/or concomitant administration of nonsteroidal anti-inflammatory drugs (NSAIDs), opioid analgesics, cephalosporins, diuretics, antimycotics, inotropics agents, and corticosteroids. The preliminary selection of covariates was performed by stepwise generalized additive model (GAM) analysis. Covariate selection was guided using likelihood ratio tests at a significance level (P value) of <0.05 for forward addition of covariates (ΔOFV > 3.84) and P value of <0.01 for backward elimination of covariates (ΔOFV > 6). Diagnostic plots and comparisons of changes in the minimum OFV between the nested models were used to evaluate the covariates and physiologically reasonable results.

Continuous covariate effects were introduced into the population model using linear, power, or exponential functions; parameters were also centered on the median value of the continuous variables in the database (allometric function). Categorical covariates were usually set to 1 for the most frequent classification and introduced to the model as described by Mould and Upton (31).