Designing Sparse-Sampling Schemes for Population PK Study of a Highly Variable Drug.

PDF of poster

Objective: Design a sparse sampling scheme for a population PK component of an efficacy trial of an anti-tumor drug.  Due to predisposition to bleeding of the study patient population, the optimal sampling scheme should incur a minimal number of blood draws per patient. The data collected will augment available PK data and allow for identification and testing of covariates for this highly variable drug.

Methods: A population PK model was developed based on the data from 7 Phase I clinical trials, wherein patients (n=68, 40% female, 60% male) were given twice-daily doses of 200, 250 or 300 mg.  A total of 727 plasma concentrations (n=7-14/patient) were available for 12 or 24hr post-dose at steady-state. Model performance was assessed by evaluation of diagnostic plots and the final model was identified.  For each sparse sampling scenario, individual patient’s PK profile was bootstrapped to generate 100 datasets. The population PK model was fit to each bootstrapped dataset and accuracy and precision of resulting parameter estimates were compared to those from the original model. All data modeling and simulations were performed using S-Plus v 6.0 and resampling using SAS 8.

Results: A steady-state one-compartment with an additive interindividual random variability and residual error model best described the data. The final model did not have any covariates due to limited diversity of covariate data (race, gender and weight). The likelihood of a successful model fit to the bootstrapped datasets was related to the number of sampling points in the design. The percentage of datasets with successful model fits ranged from 70%-95% for the different sampling scenarios.  For the designs under consideration, the scaled mean squared error (MSEs) for the population model parameters ranged from 0.0056 to 0.0171. 

Conclusions:  The ease and flexibility of proposed methodology allowed for evaluation of various sampling scenarios which were subsets of the original design.  The obtained estimates of each model fit were used for computation of standard metrics of model prediction errors and design efficiency.