Xavier Barbaut (1,2), Roger W. Jelliffe (2), Pascal Maire (1,2), Alan Schumitzky (2), Michael van Guilder (2), Bruno Charpiat (1)
(1)ADCAPT, Geriatric Hospital A. Charial, Hospices Civils de Lyon, 40 Avenue de la Table de Pierre, 69340 Francheville, France; (2)Laboratory of Applied Pharmacokinetics, School of Medicine, University of Southern California, 2250 Alcazar Street, Los Angeles, CA 90033
Schumitzky has recently developed a nonparametric expected EM algorithm that represents another approach to the method of Mallet, who introduced the nonparametric approach in population pharmacokinetics. This presentation of the new version of NPEM, NPEM2, will focus at first on the clinical use of NPEM, and secondly will show how nonparametric population modeling, giving the full probability density function (pdf) can now be linked to therapeutic drug monitoring and adaptive feedback control of drug regimens. NPEM is a nonparametric algorithm. That means that no parametric assumptions have been made about the form of the underlying population distribution, which will be entirely estimated from the population data. The final distribution is not assumed to have any special form, even if sometime it might, by coincidence, be normal, log-normal, bi-modal, multi-modal, etc…
NPEM2 uses the same NPEM algorithm as NPEM1, but a different integration method, currently computing the pdf for up to 7 parameters, using a model having an absorptive, a central (serum concentration), and a peripheral (nonserum) compartment. Oral doses can now be handled. Current parameters include Ka, Vd, Kcp, Kpc, Ke (or Ki and Ks), Cl, and F (fraction orally absorbed). With data of combined oral and IV doses and relevant serum levels, both Vd and F can be obtained. Plots of each marginal pdf are made, and 3D plots of the joint pdf of parameter pairs, providing a visual addition to the correlation and covariance matrices. Means, SD’s, modes, medians, skewness, kurtosis, and various percentile values are also given. Prior to the use of NPEM2 itself, an iterative Bayesian subroutine computes for each individual the Bayesain posterior in order to determine the probable bounds of the parameter space within which NPEM2 will run. NPEM2 uses a polynomial up to third order to describe the profile of the assay standard deviation (SDc) over its entire working range, allowing a linear or nonlinear relationship between SDc and concentrations for any level. This permits more accurate determination of the Fisher information (the credibility) of any given serum level. This part of the environmental noise is different from the other noises (in drug preparation, in time of administration, in time of sampling, of model misspecification), which will be also incorporated in the future. It is also very easily determined.
NPEM2 is part of the USC*PACK PC Clinical Programs: the physician or clinical pharmacist in charge of the therapeutic drug monitoring in a hospital will be able to use NPEM2 retrospectively on the patient data files he has prepared using the PASTRX program during routine clinical pratice. Along with NPEM2, a Multiple Model Linear Quadratic (MMLQ) controller has been developed by the USC laboratory. The MMLQ controller uses a multiple model approach which exactly incorporates Bayesian priors based on the nonparametric models. Thus, the entire joint pdf computed by the NPEM algorithm can now be incorporated as the Bayesian prior for MMLQ, permitting improved open-loop – feedback dosage strategies.
Reference: PAGE 3 (1994) Abstr 842 [www.page-meeting.org/?abstract=842]
Poster: Software Demonstration