Robert Bies (1,2), Mark E. Sale (3), Bruce G. Pollock (1,2)
(1)Department of Pharmaceutical Sciences, School of Pharmacy, (2) Department of Psychiatry, Western Psychiatric Institute and Clinic, University of Pittsburgh, Pittsburgh, PA (3)Glaxo Smith Kline, North Carolina, USA
Covariate search strategies have provided some measure of controversy with respect to the identification of significant or correlated effects. Using a dataset of 58 subjects administered IV citalopram infusion with intensive plasma concentration sampling, two covariate search strategies were undertaken. The first was a stepwise covariate search, employing both forward addition/backward removal and backward removal/forward addition directions under both First Order and First Order Conditional estimation conditions. The second was an automated covariate search strategies employing a genetic algorithm approach under the first order estimation condition. All covariate searches were carried out using NONMEM. A two-compartment model was used to describe citalopram pharmacokinetics after IV administration. The covariates age, weight and sex were evaluated on each of the parameters CL, Q, V1 and V2. The genetic algorithm approach may evaluate the search space identifying effects that are only present when in combination. Results from the binary tree approach yielded different models depending on the search direction and order. The genetic algorithm search resulted in yet a different model evaluated with a lower objective function (under first order estimation conditions). The genetic algorithm search was sensitive to the number of individual models specified per generation (individual=model). Searches with smaller numbers of individuals resulted in the identification of an optimal region of solutions. As the numbers of individuals increased a single optimal solution was detected. The model identified with the GA approach had the lowest objective function detected (D>30, Dq=+2, Dh=-2 relative to the best model found with the stepwise approach). The final model from the GA method identified two inter-individual variability terms and incorporated all three covariates (sex, weight and age) on clearance and two covariates (clearance and sex) on V2 into the model.
Reference: PAGE 12 (2003) Abstr 405 [www.page-meeting.org/?abstract=405]
Poster: poster