Detection of Outliers and Accuracy of Pharmacokinetic Parameter Estimation Using Parametric and Non Parametric Population Modeling Approaches

CACERES GUIDO P; NEELY MN; NISELMAN AV; SCHAIQUEVICH P

Salt Lake city

Congreso; XXII Congress of the Theapeutic Drug Monitoring and Clinical Toxicology society; 2013

Background: The non-parametric (NP) population algorithm in Pmetrics has recently been reported to obtain better estimates of the true individual phar-macokinetic (PK) parameters than the Pmetrics parametric (P) algorithm when outlier subjects are present.1 Here, we studied the infl uence of outlier values on the estimation of individual pharmacokinetic parameters in a real population of patients, comparing the Pmetrics NP algorithm and a second P algorithm, in the Monolix software Methods: The PK data consisted of 100 measured concentration-time points obtained from 27 neonates who received amikacin. We fit the data to a one-compartment model parameterized with kel and Vd, using the Pmetrics NPAG algorithm for the NP approach and the Monolix MCMC-SAEM algo-rithm for the P approach. The NP method detected an outlier, and each method was then used to fi t the data without this patient, and different aug-mented datasets where the outlier patient was repeated up to n = 30 times (datasets S0 ?S 30). We defi ned the %Bias for kel and Vd for each non-outlier i- patient as j thetan,i?theta0,i j /theta0,i with n . 0. We compared the median %bias of the population parameters between NP and P approaches for S1 ?S30. Results: For the original data set with 1 outlier (S 1) the mean NP and P population parameters were kelNP 0.19 h2 1,VdNP 0.50 L; kelP 0.16 h2 1, and VdP 0.44 L, respectively. However, for the outlier, the NP and P Bayes-ian posterior parameters were kelNP 0.497 h2 1,VdNP 0.025 L; kelP 0.147 h2 1, and VdP 0.448 L, respectively. Despite the monotonically increase in median %bias with both the NP and P methods, as n increased to 30, the median % bias was consistently lower with the NP approach, showing less in fluence of outliers on individual parameter estimation than the P method. For S30, the median %bias for kel and Vd were 18.9% and 13.6% for NP and 34.6% and 15.7% for P, respectively. Co nc lu si o ns: Compared to at least 2 parametric algorithms, the non-parametric algorithm as implemented in Pmetrics more accurately detects outliers and minimizes biases on parameter estimations induced by outliers