Back to the Top
I'm not sure which question/response prompted this but I figured I'd
toss my two
cents in.
My hospital has been using USCpack for several years and it had been VERY
useful. Interesting enough, the best part about MAP Baysean modeling
is not the
predictive tool but rather the fact that data points are conserved.
A picture of
what is happening with the patient can be build that is dynamic rather than
static. In using this software (or similar programs) what I've been
most excited
about is that all data points are useable. Even misdrawn serum levels can be
used to build a dynamic picture of the patient.
Being able to get a 'feel' of how your patient is handling a
drug, even when
you know that the information is not precise is incredibly valuable.
Bob
Back to the Top
Hi Bob
You comment that one of the benefits of MAP (whether USCPack
or otherwise) is that you can fit a model to all datapoints
simultaneously hence conserving all data.
> but rather the fact that data
> points are conserved.
> A picture of
> what is happening with the patient can be build
> that is dynamic rather than
> static. In using this software (or similar
> programs) what I've been
> most excited
> about is that all data points are useable.
I would argue that this model is static rather than dynamic
(in an english rather than engineering sense) since it does
not allow for a patient to change (independent of their
covariates) over a course of treatment. Eg in the case of
antibiotics it is presumed the patient will get better
during treatment which is often accompanied by a change in
their parameter values without necessarily a change in their
"predictive" covariates. How do you account for this?
Regards
Steve
=================
Stephen Duffull
School of Pharmacy
University of Queensland
Brisbane, QLD 4072
Australia
Ph +61 7 3365 8808
Fax +61 7 3365 1688
Back to the Top
[The email system misbehaved a little today...the PharmPK listserv
members may have received multiple messages...or delayed
messages...or missed messages (see the archive if this is what
happened) I think its working properly again, I hope ;-) - db]
I think the underlying message is that a physician asked to track too many
patients is better off with a tool that reports data, accurate or
inaccurate, about his patients rather none whatsoever.
This is a statement of the state of patient administration under which some
physicians must attempt to treat patients, absurd as it is on the face.
Hopefully, this will change as the law suits against the big HMOs for
allowing patients to be harmed in order to save moneys take off in the courts.
Daro Gross
Back to the Top
Steve,
I'm not sure how you see this as static. At each point that you
print a current analysis I suppose that you could consider that point
static. Still, with each data point you enter, the picture changes and
evolves.
The result is a very useful tool that models how an individual
patient handles a drug over time. Given a choice, I'll take Bayesean
modeling over linear regression any day you care to name. Given the
fact that aminoglycoside induced nephrotoxicity is not an issue at my
institution and that we provide very good, very therapeutic
aminoglycoside dosing, I'll stick with it for now.
Basically it just plain works.
Bob
Back to the Top
[Two replies - db]
From: "Stephen Duffull"
To:
Subject: RE: PharmPK Re: Population Modeling
Date: Thu, 24 Feb 2000 15:43:33 +1000
X-Priority: 3 (Normal)
Importance: Normal
Bob
> I'm not sure how you see this as static. At
> each point that you
> print a current analysis I suppose that you could
> consider that point
> static. Still, with each data point you enter,
> the picture changes and
> evolves.
I'm sure the picture does evolve - but in essence the MAP
method gives equal weighting to all data points (whether
sampled an hour ago or a week ago). Hence the solution to
the Bayesian problem will be an average over the whole
course. What I am asking is whether we can expect a patient
to remain stable over a course?
> The result is a very useful tool that models
> how an individual
> patient handles a drug over time. Given a
> choice, I'll take Bayesean
> modeling over linear regression any day you care
> to name.
I am not disputing the value of MAP over other tools for
dose individualisation - however I do suggest that there
maybe caveates within the MAP process which need to be
considered from time to time.
Regards
Steve
=================
Stephen Duffull
School of Pharmacy
University of Queensland
Brisbane, QLD 4072
Australia
Ph +61 7 3365 8808
Fax +61 7 3365 1688
---
X-Sender: mentor.-at-.hardlink.com
Date: Thu, 24 Feb 2000 04:44:46 -0500
To: PharmPK.-a-.boomer.org,
Multiple recipients of PharmPK - Sent by
From: Daro Gross
Subject: Re: Population Modeling
Bob:
I presume that, in simple English, it is more useful to you to identify
patients that respond differently than other patients. Patients whose
response differs from the expected is not useful information, hence your
choice of analysis. You have discarded a tool for identifying errors or
incorrect assumptions of any kind. Are you sure that is such a wise idea,
even if your institution is of the highest quality?
Daro
Back to the Top
Daro,
1. It is very useful for me to identify patients that outside the norm.
Those are the ones that potentially have the most problems.
2. In a healthy 35 y/o patient I can safely dose aminoglycosides using
anything from a mg/kg recommendation all the way through the most
sophisticated modeling program you choose. They aren't the ones that are
problems. It's the 85 y/o debilitated patient that I really need to be
concerned with. The one who doesn't fit the norm.
3. While I'm in no way a statistics expert, my understanding of Bayes
Theorem is that it provides a model that responds to new data. This is a
very big plus.
4. After dosing several hundred patients with aminoglycosides over the
past couple of decades, using everything from natural log tables and a piece
of paper to programs that use Baysean modeling, I'll stick with what I've got
until something better comes along. Results are what's important.
5. The bottom line for where I practice is getting the patient well and
causing as little harm along the way as possible. The elegance of research
is a bit beyond me. I'll stick with afebrile patients with intact kidneys
along with any other analogous situations I can help provide.
6. I'll climb off my soap box now
Bob
Back to the Top
I'm sure the picture does evolve - but in essence the MAP
method gives equal weighting to all data points (whether
sampled an hour ago or a week ago). Hence the solution to
the Bayesian problem will be an average over the whole
course. What I am asking is whether we can expect a patient
to remain stable over a course?
Regards
Steve
=================
Stephen Duffull
School of Pharmacy
University of Queensland
Brisbane, QLD 4072
Australia
Ph +61 7 3365 8808
Fax +61 7 3365 1688
Back to the Top
[Two replies - db]
Comments: Authenticated sender is
=46rom: "Guillermo Bramuglia"
To: PharmPK.aaa.boomer.org
Date: Fri, 3 Mar 2000 14:11:42 +0000
Subject: Re: Population Modeling
X-Confirm-Reading-To: "Guillermo Bramuglia"
X-pmrqc: 1
Priority: normal
some programs allow to weight points using an error function (such as
USC pack). This is a way to differentiate the points to fit taking
information about the analitical technique, for example. The problem
is (I think), how to include different kinds of information about the
patient that can be take into account over a course, to fit and
estimate the best parameters.
regards,
Guillermo
Dr. Guillermo Bramuglia
C=E1tedra de Farmacolog=EDa,
=46acultad de Farmacia y Bioqu=EDmica,
Universidad de Buenos Aires.
---
Date: Fri, 03 Mar 2000 20:17:38 -0500
=46rom: Brennan
X-Accept-Language: en
To: PharmPK.-a-.boomer.org
Subject: Re: Population Modeling
But as each data point is weighed, the overall picture of what is
happening also changes. All of this must be in conjunction with the
clinical assessment of each patient of course.
Let's put it this way. If anyone has a method of analysis that does a
better job in a non-research, on the floor, "I need the answer now" type
of setting, I'd love to hear about it.
=20 Bob
Back to the Top
Guillermo,
Your comments speak very well to the clinical setting. Indeed you point
out the problem itself. How best to use the analytical tools available
versus how best to treat the patient.
The number of variables involve in any given patient is huge and cannot
be modeled with the precision needed for PK/PD solutions designed for that
individual patient. There are far to many.
By defining populations, one can narrow the variables down a bit. By
sampling and remodeling, one can start to see the differences in each
patient. It is here that an optimal therapeutic regimen can be outlined.
Is it the best way? Hopefully not. Is it valuable? Most definitely!
Bob Brennan
Back to the Top
Guillermo,
Your comments speak very well to the clinical setting. Indeed you
point
out the problem itself. How best to use the analytical tools available
versus how best to treat the patient.
The number of variables involve in any given patient is huge and
cannot
be modeled with the precision needed for PK/PD solutions designed for that
individual patient. There are far to many.
By defining populations, one can narrow the variables down a bit. By
sampling and remodeling, one can start to see the differences in each
patient. It is here that an optimal therapeutic regimen can be outlined.
Is it the best way? Hopefully not. Is it valuable? Most definitely!
Bob Brennan---
Back to the Top
Hi Bob and Steve Duffull:
About MAP and static versus dynamic. What we have done with
the USC*PACK
software is to parameterize the models so they can change with changing
renal function and body weight. As the patient's renal function changes, so
does his kel, for example, as it is described as being a nonrenal intercept
value times a slope value (increment of Kel per unit of CCr) times the
patient's CCr at any time. That is why we developed the method for
estimating CCr from changing, and not just stable, serum creatinine values.
The model is able to follow and track the changes in the status of the
patient from dose to dose and from day to day.
One can then use MAP Bayesian fitting procedures to fit the
model to the
data of the populatioon prior model and the patient's data of dosage, CCr,
body weight, and serum concentrations, and come up with a model that can
track the behavior of the drug in the patient over time.
The interesting thing is the point you raise about what to do when the
patient's parameters change independently of the covariates. How does one
detect these changes? This is a sticker. All the procedures out there at
present that I am aware of assume that the patient is fixed and unchanging
in his/her parameter values during the fitting procedure. Even with
sequential Bayesian approaches that I know of, this is the case. For
example, if one has a cluster of 4 serum levels to fit, whether they are
fitted sequentially or all together, while the sequential method shows
changes in the Bayesian posterior during the fitting and addition of each
new data point, I am told that one has still not departed from the
hypothesis that the true patient has fixed and unchanging parameter values,
and that one has not learned anything new from the point of view of optimal
dosing than from fitting them all 4 together.
However, an interacting multiple model (IMM) approach to developing
Bayesian posteriors that can detect such a change in parameter values over
time has been developed by Dr. David Bayard in our lab. When presented with
a simulated patient in which the changes in the parameter values occur at a
stated time in the regimen, the sequential IMM approach tracked the
behavior of the simulated patient with only about half the integrated error
of the sequential MAP Bayesian or the sequential multiple model Bayesian
methods, and was able to do a reasonable job of detecting the change. This
work was presented at the SCS meetings in San Diego in January,
(Proceedings of the Society for Computer Simulation International, Health
Sciences Simulation, 2000, pp 75-83). It is being incorporated into the
new Windows version of the Multiple Model USC*PACK software which we hope
to have available for clinical use sometime this year.
Very best regards,
Roger Jelliffe
Roger W. Jelliffe, M.D. Professor of Medicine, USC
USC Laboratory of Applied Pharmacokinetics
2250 Alcazar St, Los Angeles CA 90033, USA
Phone (323)442-1300, fax (323)442-1302, email= jelliffe.-at-.hsc.usc.edu
Our web site= http://www.usc.edu/hsc/lab_apk
**************
Back to the Top
Dear Steve and Guillermo:
About the MAP Bayesian process and stability. It is nice to link PK
parameters to clinical descriptors as much as possible, and to parameterize
the parameters in terms of these descriptors or covariates as much as
possible, so that when the patient changes, and the descriptors change, one
can still follow the changing behavior of the patient. In this, there is
still the built in hypothesis that the "patient" being fitted is fixed and
unvarying thrung the period of the data analysis, even is the MAP Bayesian
analysis is being done sequentially, as each new data point becomes available.
To overcome this, and to be able to detect changing parameter
values which
are NOT accompnaied by changes in descriptors or covariates, Dave Bayard in
our lab has implemented an Interacting Multiple Model (IMM) approach to
obtaining discrete Bayesien posteriors in which the patient's parameter
values actually may jump from one set of values in the nonparametric
population discrete joint density to another. This approach comes from the
aerospace community, and has worked well in tracking a simulated patient
whose parameter values changed, with only about half the total error if the
MAP or multiple model approach. This work was presented at the CSC meetings
in San Diego in January. I am enclosing a copy of an abstract about it
which we have submitted for consideration for presentation at the PAGE
meeting in June. It appears to be a distinct improvement over the MAP
approach, and it builds on the strengths of nonparametric populationmodels
in general.
AN INTERACTIVE MULTIPLE MODEL (IMM) ESTIMATION APPROACH
TO UPDATING POSTERIORS IN PHARMOCOKINETIC MODELS HAVING CHANGING PARAMETER
VALUES
Roger W. Jelliffe1 and David S. Bayard2,
1 Laboratory of Applied Pharmacokinetics, University of Southern
California, School of Medicine, Los Angeles CA, 90033
2 Senior Research Scientist, Jet Propulsion Laboratory, Pasadena CA, 91109
This report considers updating Bayesian posterior densities for
pharmacokinetic models having changing parameter values. The prior
probability is assumed to be a discrete joint density. Parameter changes
are modeled as "jumps" from one model support point to another within the
same discrete density. Given such discrete priors, the multiple model (MM)
estimation approach provides an exact analytical solution for updating the
Bayesian posteriors. Our laboratory's earlier studies showed that the MM
estimator works well in pharmacokinetic applications where the patient's
parameters are unknown but constant.
Unfortunately, the MM estimator works less well where the patient's
parameter values vary. The IMM algorithm has emerged as an effective method
in the literature for tracking changing parameter values, and is used by
the aerospace community for tracking maneuvering targets. We implemented
the IMM sequential Bayesian algorithm in pharmacokinetic software. Its
performance was compared with the MM and MAP sequential Bayesian estimation
methods, which are used in pharmacokinetic applications where parameter
values do not change, using both simulated and real clinical data for the
drug Tobramycin.
In a simulated therapeutic scenario of changing parameter values taking
place at a stated time, the IMM approach tracked the behavior of the
simulated patient with about half the integrated total error found with the
MM and MAP methods. Further, in examining a real patient's data, in which
parameter values appeared to change significantly during therapy, the IMM
approach tracked the patient's data quite acceptably, and much better than
the MAP Bayesian approach.
While the present paper focuses on estimation, the authors' main motivation
for studying this problem is to eventually develop better controllers,
i.e., to extend the present Multiple Model active control approach to
designing dosage regimens for unstable patients with varying parameter values.
Supported in part by NIH grants LM 05401 and RR 11526
Very best regards,
Roger Jelliffe
Roger W. Jelliffe, M.D. Professor of Medicine, USC
USC Laboratory of Applied Pharmacokinetics
2250 Alcazar St, Los Angeles CA 90033, USA
Phone (323)442-1300, fax (323)442-1302, email= jelliffe.aaa.hsc.usc.edu
Our web site= http://www.usc.edu/hsc/lab_apk
********************************************************************
PharmPK Discussion List Archive Index page
Copyright 1995-2010 David W. A. Bourne (david@boomer.org)