Back to the Top
Dear all:
I am a masters student working on a project attempting to do population
pharmacokinetics on some PK data we have on antiretrovirals in HIV
patients, and I am totally new to population PK. We have the WinNonMix
Professional 2.0.1 program (We have both the WinNonlin and WinNonmix),
and I am having trouble understanding the use of WinNonMix to model our
data. I have gone through the user gudies and references but they
don't help at all since they only provide instructions as to where to
click and how to enter values, but no explanations at all regarding
fixed effects and random effects (for example). I would appreciate any
help/hints/pointers in using and understanding this program.
We basially have a bunch of concentration vs. time PK data, and for
example:
1) we have to enter the mixed effect equation(s) for V/F, K10 and K01.
From the concentration vs. time PK profile data we have, how do we
know what kind of random effect function we should use? (e.g. log,
exponential, linear etc.), and what kind of variance function (e.g.
power, linear etc.) to input for error structure specification? Do we
have to obtain these functions through other ways using our conc vs.
time data, or just guess?
2.) I know this is a program-specific question...but we are asked to
input initial estimate for the paramters (V/F, K10, K01), yet the
program does not specify the units for these parameters (L? mL? h-1?),
and I couldn't find instructions anywhere in the 'help' section. Has
anyone used this program and can give me some insight? My 'modelling'
obviously doesn't work despite numerous 'guessing' in the units, since
all my 'predicted' values for the parameters came out as zero.
I would appreciate any help and suggestions on reading materials and
references regarding this topic. Will I need specialized training
(e.g. workshops, classes) for me to know how to do population
modelling?
Thanks!
Lillian Ting
MSc. candidate, Pharmaceutical Sciences
University of British Columbia
Back to the Top
Dear Ms. Ting:
You might ask whether or not your software has the guarantee
that
the more subjects you study, the closer the results get to the true
ones.
No currently available software for parametric modeling has this
guarantee.
Also, that about the precision of parameter estimates?
If you go to our web site, www.lapk.org, and click on new
advances
in population modeling, you will see that Dr. Robert Leary has shown
that
methods which use the FO or FOCE approximation to compute the likelihood
function, such as the ones you describe, do not have the guarantee of
consistent behavior, nor do they have the best statistical efficiency
(precision of parameter estimation). However, nonparametric methods
such
as the NPML of Alain Mallet, and the NPEM, NPAG, and NPOD by Schumitzky
and
by Leary, from our lab, easily compute the likelihood function exactly,
and
therefore do in fact have the proven property of consistency, and much
better statistical efficiency than methods that use approximate
likelihoods.
Because of this, if you are interested in population PK/PD
modeling, you might consider using proven consistent and precise
methods.
In addition, the NP models lend themselves best to developing dosage
regimens that are maximally precise, as they permit the calculation of
the
expected precision with which any regimen will hit a desired therapeutic
target, using multiple model (MM) dosage design. Information on how to
get
the software is also available.
Very best regards,
Roger Jelliffe
Roger W. Jelliffe, M.D. Professor of Medicine,
Division of Geriatric Medicine,
Laboratory of Applied Pharmacokinetics,
USC Keck School of Medicine
2250 Alcazar St, Los Angeles CA 90033, USA
Phone (323)442-1300, fax (323)442-1302, email= jelliffe.aaa.usc.edu
Our web site= http://www.lapk.org
Back to the Top
Lillian,
It seems like you need some basic training in the terminology and
methods of population modelling. This should then help you figure out
how to use WinNonMix that are available.
The random effects for parameters and residual error are often chosen
to arise from a log normal distribution. This assumption is usually
quite robust. You may need an additional additive random effect on the
concentration residual error. If this advice does not mean anything to
you then you really do need some basic training. It shouldn't take long
if you can meet up with someone familiar with this area.
With regard to units -- its pretty much up to you and not the program.
If doses are in mg and concentrations in mg/L and time in hours then
volumes will be L, clearances in L/h, rate constants h-1 (K10 etc).
My personal advice is not to bother with WinNonMix. It might seem
easier to use than other programs initially because the interface is
like WinNonLin but I have found that it quickly becomes hard or
impossible to do things you might want to do that are much easier with
other programs (e.g. NONMEM).
Nick
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New
Zealand
email:n.holford.aaa.auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
Back to the Top
Hi Lillian,
Take a look at these articles to start with-
1. Sheiner LB and Grasela TH: An introduction to mixed effect modeling:
Concepts definitions and justification J Pharmacokin Biopharm
19[3]:11S-24S,
1991
2. Grasela TH and Sheiner LB: Pharmacostatistical modeling for
observational
data J Pharmacokin Biopharm 19[3]:25S-36S, 1991
Then more detailed theory can be learned from statistical literature
&/or
books. My personal preferences are (examples provided deal with PK/PD
models):
1. Model building for Nonlinear Mixed-effects models
http://www.pharmastatsci.com/download/model%20building.pdf
2. Mixed effects modeling in S and S-Plus by Pinheiro and Bates 2000
ISBN
0-387-98957-9
3. fitting PK models with NLME
http://www.insightful.com/events/2002uc/speakers/pinheiro.pdf
And I agree with Nick. There are softwares around like NONMEM that are
very
flexible and powerful.
Thanks,
Pravin
Pravin Jadhav
Graduate student
Department of Pharmaceutics
MCV/VCU
Back to the Top
Dear Lillian:
About population modeling. There is much more and better stuff
around than just the present NONMEM. Ask whether NONMEM has guaranteed
mathematical consistency. It it really true that the more patients you
study in a population, the closer the parameter estimates get to the
truth?
No guarantee. What about the efficiency (precision) of the parameter
estimates? Much better with methods that use exact likelihoods.
Nonparametric population (NP) approaches use exact likelihoods, which
are
usually not reported in studies using NONMEM or similar approaches which
use approximations (FO, FOCE) to compute the likelihood. NP methods are
consistent, and have much better efficiency (more precise parameter
estimates).
Many populations have genetically determined subpopulations
such
as fast and slow metabolizers. These will not be detected by parametric
methods without further help (covariates, etc.). NP methods estimate the
entire parameter distributions, not simply means and covariances.
However,
since NP methods are consistent, so are their means and covariances.
Better
estimates than with NONMEM of other methods that use FO or FOCE
approximations of the likelihood.
Further, what will you do with the results you get? Do you just
want to report them, or do you want to develop dosage regimens to
achieve
selected therapeutic target goals? With parametric methods, there is no
way
to estimate in advance the precision with which you will be able to hit
a
therapeutic target. There is only 1 model, which is based on the
estimated
central tendencies of the parameter distributions. The full shape of the
distribution cannot be considered unless it is assumed to be symmetrical
about the central tendency.
With NP pop models, though, you have many support points
reflecting the entire most likely distribution of the parameter values
in
the population. You now have many support points to make predictions
from.
Each of these predictions can be weighted by its probability. Because of
this, any candidate dosage regimen can be evaluated by its ability to
hit
the target, by computing the weighted squared error of the regimen's
failure to hit the target. Now, one can find the specific regimen which
minimizes that weighted squared error. This is the maximally precise,
multiple model (MM) dosage regimen.
We now have software to make such NP population models (small
and
large, linear and nonlinear), and to examine relationships between
parameters and covariates. We also have clinical software for MM dosage
regimens for patients having 2 compartment linear models. They can track
the behavior of the drug well, even in acutely ill, unstable patients,
and
can develop maximally precise dosage regimens for them. They are in beta
phase release from our web site www.lapk.org/beta. Other demonstration
software can be obtained from www.lapk.org, as well as information on
obtaining the USC*PACK software. You can also go to teaching topics on
the
web site, and get much more information about all this, as well as
references.
Very best regards,
Roger Jelliffe
Roger W. Jelliffe, M.D. Professor of Medicine,
Division of Geriatric Medicine,
Laboratory of Applied Pharmacokinetics,
USC Keck School of Medicine
2250 Alcazar St, Los Angeles CA 90033, USA
Phone (323)442-1300, fax (323)442-1302, email= jelliffe.aaa.usc.edu
Our web site= http://www.lapk.org
PharmPK Discussion List Archive Index page
Copyright 1995-2010 David W. A. Bourne (david@boomer.org)