A comparison of a genetic algorithm based automated search algorithm to standard stepwise approach for population pharmacokinetics using NONMEM.
Robert R. Bies, Jeff A. Florian, Bruce G Pollock, Kristin L Bigos, Marci Chew, Yuyan Jin, Yan Feng, Kim C Coley, Lon Schneider, Matt Muldoon, Steven Manuck, Mark E. Sale.
University of Pittsburgh (RRB, JAF, YJ, MM, SM, KCC), CAMH (BGP, RRB), NIMH (KB), Pfizer (MC), BMS (YF), USC (LS), Next Level Solutions (MES).
Objectives: To compare, using the Akaike Information Criteria (AIC), final population PK models selected using an automated search algorithm versus a standard stepwise approach.
Methods: Five population PK analyses were available for this comparison comprising the PK of citalopram, perphenazine, olanzapine, quetiapine, and risperidone. All five available analyses were repeated using the automated approach. All analyses were performed blindly (i.e. the analyst was not aware of the outcome of the original analysis). The automated search algorithm search space was limited in scope to those elements that were evaluated using the standard stepwise approach. In all cases except one (risperidone), this included evaluating a one versus two compartment PK model structure, the form of the inter-individual variability (exponential, proportional, additive), the form of the residual error (additive, proportional, combined), the presence of covariate relationships, the mathematical structure of these covariate relationships and multiple sets of initial estimates. In the case of risperidone, a mixture model option was included in the search space for the automated method, as one was assessed during the stepwise search. The models were all tested using NONMEM VI using the genetic algorithm (Bies 2006) and the FOCE Inter estimation option.
Results: 5/5 models had lower AIC values (p=0.03125) as selected by the genetic algorithm approach versus the stepwise approach. The geometric and arithmetic means of this reduction were 217.7 and 465.8 with a range of 5 to 1017 points.
Conclusions: In this test set of population PK examples, the genetic algorithm based automated search was consistently able to find models with lower AIC values than the stepwise approach.
References:
[1] Bies RR, Muldoon MF, Pollock BG, Manuck S, Smith G, Sale ME. A genetic algorithm-based, hybrid machine learning approach to model selection. J Pharmacokinet Pharmacodyn. 2006 Apr;33(2):195-221. Epub 2006 Mar 28.
