**From rs@chdr.leidenuniv.nl Tue Sep 24 00:50:01 1996
**

I'm running into the following problem. I'd like to do PK/PD modelling for a benzodiazepine that has been administered orally. However kinetics are rather irregular (deviations from simple first or zero-order absorption) making a good description/interpolation of the kinetic profile impossible

using standard pharmacokinetic models. After glancing over the nice work of Kyungsoo Park et al I wondered if something like constrained longitudinal splines could work in this case as well. I feel I may have to resort to an effect compartment so I need some sort of continuous interpolation to feed into my compartment. I was hoping this is all clear cut for someone out there! If not, I welcome any suggestions.

Regards,

Rik Schoemaker,

CHDR, Leiden, NL

****

**From kyungsoo Tue Sep 24 10:06:09 1996
**

First of all, the constrained longitudinal spline (CLS) method is for dealing with population PK data. For the benzodiazepine under consideration, if you have PK data as well as PD data, I think,

you can use CLS to first fit PK data, thus obtaining the estimated kinetic profile for each individual. Then, substituting the estimated kinetic profiles into the PK portion of your selected PK/PD model while treating them as constants, you will be able to fit your PD data.

If there are only PD data, other type of spline model may be used to describe the kinetic profile.

Hope this helps.

Kyungsoo Park

UCSF

****

**From kgkowa@searle.monsanto.com Tue Sep 24 12:13:48 1996
**

Rik,

You might want to consider the semicompartmental modeling approach that I have been developing. The semicompartmental approach is a solution to Sheiner's effect-site link model based on a noncompartmental model for Cp (piecewise linear or log-linear model) and is easily implemented in standard nonlinear regression packages such as NONLIN, NONMEM, and SAS NLIN. I was motivated to develop this approach precisely for the reason you indicated in your message, ie., when the kinetic profile is not accurately described by standard compartmental pharmacokinetic models. The catch with my approach is that you need to have enough times points at properly spaced intervals such that the AUC can be accurately estimated by linear and/or log-linear trapezoidal rule calculations. Here is the reference for this approach:

Kowalski, K.G. and Karim, A. A Semicompartmental Modeling Approach for Pharmacodynamic Data Assessment. J. Pharmacokin. & Biopharm., 23:307-322 (1995).

If you decide to try the semicompartmental modeling approach, let me know if I can be of further assistance.

Ken Kowalski

G.D. Searle

Skokie, IL

****

**From n.holford@auckland.ac.nz Tue Sep 24 14:39:53 1996
**

I am bit puzzled by this reply. What do you mean by a "kinetic profile"? If you mean predicted concs at the times of each of effect observation this is not much help in using a parametric model for an effect compartment although it might be used for one of the semi-paramtric loop collapsing methods for estimating the equilibration half-time.

I presume you do not mean a set of PK parameters such as CL, V, KA, because CLS does not deal with this kind of parameterisation. Or perhaps you mean that there are a set of parameters from CLS that can be used? If this is the case then an example control stream would be helpful.

If one can use CLS to interpolate the predicted PK concs then it should be possible to use these with a differential equation defined model to predict effect cpt concs or use a physiological delay model (aka indirect response model). I have used a cubic splines with MKMODEL to interpolate individual PK data and then used a DE model to estimate effect cpt model parameters. It would be nice to see how to do this using CLS in the NONMEM environment.

--

Nick Holford, Dept Pharmacology & Clinical Pharmacology

University of Auckland, Private Bag 92019, Auckland, New Zealand

email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556

http://www.phm.auckland.ac.nz/Staff/NHolford/nholford.html

****

**From kyungsoo Tue Sep 24 17:40:54 1996
**

When I replied to Rik Schoemaker's question, I was thinking of two different cases. The first case is where not only the effect measurements but the concentration measurements are also available

at each time of effect observations. The second case is where only the effect measurements are available.

Regarding the issues addressed in the above message, here is what I thought, assuming PK data are also available (i.e., the second case).

By "kinetic profile", I mean predicted concs at the times of each of effect observation.

By "estimated", I mean, from PK data, estimating THETA and ETA which parameterize CLS where THETA is a vector of population parameters common to all individuals (but does not mean PK parameters such as CL, V, etc), and ETA is a vector of inidividual-specific random effects.

Thus, given THETA and ETA, the individual's concentration prediction at each effect observation time would be obtained as the CLS value at that time.

As an example, suppose one wants to do PK/PD modeling given both PK and PD data, using CLS for PK and an effect compartment for PD. If so, in the first step, from PK data and CLS, one would obtain the individual concentration predictions calculated at each effect observation time, and save them under a new data item, say, CP. Then, in the second step, from PD data and an effect compartment, one would use the NONMEM control stream, whose fragment looks like the following, where DV denotes the effect observation.

$INPUT ID TIME DV ... CP

$PK

KEO=THETA(1)

INPUT=CP

$DES

DADT(1)=KEO*(INPUT-A(1))

Hope this helps.

Kyungsoo Park

UCSF

****

**From kyungsoo Tue Sep 24 18:32:53 1996
**

In my previous email, I wrote:

"The first case is where not only the effect measurements but the concentration measurements are also available at each time of effect observations".

But the measurement times for the concentrations would NOT need to be the same as those for the effects. All I meant was that concentration measurements are AVAILABLE. Sorry for the confusion.

Kyungsoo Park

****

**From LIAO@cosmos.prius.jnj.com Wed Sep 25 08:48:50 1996
**

Kyungsoo:

You may recall that I had asked if you can use the CLS in $DES block couple weeks ago. I am not sure the NONMEM example you gave will work. Basically, this is a hybrid model (a mixture of integrated and differential equations). In order to solve these equations, we need to define Cp(t) at each step of numerical integration of DF. Usually this step is much smaller than the PK/PD sampling interval. Can your CLS provide this Cp(t) in $DES? If you have CLS estimate prior to $DES, the Cp(t) is only defined at each observed time point.

****

**From n.holford@auckland.ac.nz Wed Sep 25 16:06:15 1996
**

Sam (and interested nmusers)

>

> Nick:

>

> I am interested in your cubic spline interpolation for PK/PD

> modeling. Do you mind provide your NONMEM control stream file to

> the Stanford's NONMEM repository ? (Adress:

> ftp://pkpd.icon.palo-alto.med.va.gov/public/) Thanks!

>

> I had asked Kyungsoo if it is possible to use CLS in $DES, he said

> no. There must be some way to implement this.

>

> Sam Liao

>

The cubic spline interpolation of PK for use with an effect compart model is implemented in MKMODEL. It is not a part of NONMEM.

I have not yet heard from Kyungsoo about using CLS in $DES although I posted this question yesterday to nmusers. Did you see the question?

--

Nick Holford, Dept Pharmacology & Clinical Pharmacology

University of Auckland, Private Bag 92019, Auckland, New Zealand

email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556

http://www.phm.auckland.ac.nz/Staff/NHolford/nholford.html

****

**From n.holford@auckland.ac.nz Wed Sep 25 16:18:40 1996
**

> As an example, suppose one wants to do PK/PD modeling given both PK and

> PD data, using CLS for PK and an effect compartment for PD. If so, in

> the first step, from PK data and CLS, one would obtain the individual

> concentration predictions calculated at each effect observation time,

> and save them under a new data item, say, CP. Then, in the second step,

> from PD data and an effect compartment, one would use the NONMEM control

> stream, whose fragment looks like the following, where DV denotes

> the effect observation.

>

>

> $INPUT ID TIME DV ... CP

> $PK

> KEO=THETA(1)

> INPUT=CP

> $DES

> DADT(1)=KEO*(INPUT-A(1))

>

>

> Hope this helps.

I am afraid not. The code above will not work properly because it assumes a fixed value for CP during the integration step

One way to deal with this is to save individual estimates (posthoc) of the THETAs used by CLS in the output of the CLS run. Then read these values in as data. The values are then passed to CLS so that CLS can predict A(1) in the DES block. That way the DE solver can know A(1) at each time point that the DES subroutine is called. This can be at any time in the interval between two effect observation times. The question I have for you is can CLS be used in this way? Will PREDPP work if you embed the verbatim code that calls CLS in the DES block?

--

Nick Holford, Dept Pharmacology & Clinical Pharmacology

University of Auckland, Private Bag 92019, Auckland, New Zealand

email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556

http://www.phm.auckland.ac.nz/Staff/NHolford/nholford.html

****

**From kyungsoo Wed Sep 25 17:45:18 1996
**

I am sorry for the wrong code. As you pointed out, the code above is not right. I was not careful enough.

When I said it is not possible to use CLS in $DES for PK/PD modeling, I meant it is not possible to use the current CLS code in $DES. In order to be used in $DES, the code would need to be a bit changed.

When I first wrote the CLS code, I was mainly thinking of using it in $PRED, exclusively for dealing with PK data. But now I realized that it could have more applications if it can be successfully implemented in $DES.

Right now, I am working on modifying the current CLS code to see if it can also be used in $DES. I will let you know when it is done.

Best regards,

Kyungsoo Park

****

**From rs@chdr.leidenuniv.nl Wed Sep 25 22:58:41 1996
**

Thanks everyone for the many extremely helpful comments (how I just love this user group!)

I've been playing with Lew Sheiner's "connecting the dots" and it seems to work; each (effect) measurement is accompanied by a slope and intercept describing the straight line that connects adjacent concentration measurements (datafile constructed outside NONMEM). This looks somewhat like this:

$PROB linear concentration interpolation vs effect

$INPUT ID OCC ET TIME CONC EFF=DV SLO INT MDV

$DATA DATA.NM

$SUBROUTINES ADVAN6 TRANS1 TOL=4

$MODEL COMP=(EFFECT,DEFOBS)

$PK

TKEO = THETA(1)*EXP(ETA(1))

KEO = 0.693/TKEO

E0 = THETA(2)*EXP(ETA(2))

SLOPE = THETA(3)*EXP(ETA(3))

B1 = SLO

A1 = INT

$DES

CPE=A1+(B1*T)

DADT(1) = KEO*(CPE-A(1))

$ERROR

CP = A1 + (B1*TIME)

CE = F

PREDI = E0 - SLOPE*F

Y=PREDI+EPS(1)

$THETA (0,1) (200,450,600) 1

$OMEGA 1 1 1

$SIGMA 1000

$EST SIGDIGITS=3 PRINT=1 MAXEVAL 9999 NOABORT POSTHOC METHOD=0

$COV

$TABLE ID OCC TIME CP CONC CE SPV PREDI

FILE=RESULT.ASC NOHEADER NOPRINT

with the datafile DATA.NM:

1 2 -10 0 .00 322.1 .00000 .00000 0

1 2 -5 0 . . .00000 .00000 1

1 2 30 30 .00 . .38000 -11.40000 1

1 2 31 33 . 329.8 .38000 -11.40000 0

1 2 46 48 . 316.2 .38000 -11.40000 0

1 2 60 60 11.40 . .19000 .00000 1

1 2 61 63 . 302.7 .19000 .00000 0

1 2 76 79 . 301.1 .19000 .00000 0

1 2 90 90 17.10 . .10333 7.80000 1

1 2 91 92 . 305.4 .10333 7.80000 0

etc

1 2 480 480 27.30 . -.00760 30.95000 1

1 2 481 486 . 364.7 -.00760 30.95000 0

1 2 1440 1440 20.00 . -.00760 30.95000 1

2 1 -10 0 .00 458.1 3.65000 .00000 0

2 1 -5 0 . . 3.65000 .00000 1

etc

However, this approach requires a differential equation while Ken Kowalski's solution seems to provide an explicit expression for the effect compartment concentration without the need to calculate the interpolating lines. It is in a recursive form however (effect compartment concentration at time t as a function of effect compartment concentration at time t-1) and I have no idea how to

implement this in NONMEM! Any clues?

Rik Schoemaker

CHDR, Leiden, NL

****

**From R.Port@dkfz-heidelberg.de Thu Sep 26 01:05:45 1996
**

Rik Schoemaker's message of Sept 26

Hi Rik,

I think for calculating the effect site concentration at time t as a function of the effect site concentration at time (t - 1) it's enough to store the current value of the effect site concentration at the end of $PRED as a variable which, at the next time point, is invoked at the beginning of $PRED. E.g.:

$PRED

IF (TIME.EQ.0) Celast = 0 ; or some other initial value

Cenew = ... Celast ... (function of Celast and the PK parameters)

effect = ... (function of Cenew and more parameters

Y = ... effect ... EPS( )

Celast = Cenew

This way, Celast should have the value of the effect site concentration at time (t - 1) when PRED is invoked at the next time point and, thus, could be used for calculating Ce at time t.

Best wishes! Ruedi

-------------------------------------------------------------------------------

R.E. Port, Dept. 0420, German Cancer Research Center

P.O. Box 10 19 49, D-69009 Heidelberg

phone: x49-6221 42-3385

-3347

fax: -3346

e-mail: r.port@dkfz-heidelberg.de

***

**From lewis Thu Sep 26 10:45:14 1996
**

To all:

It *is* possible to use plines in $DES. I have some coding for this as part of my lecture notes in my Advanced PK course, and if I find the time, I will try to make this available.

But, this is not necessary, as I wrote to Sam yesterday. A linear spline is adequate to represent Cp.

In that case, one adds 2 columns to the data records. In the first one records the slope, and in the

second, the intercept for the line connecting the bracketing observed Cp values.

To be precise, imagine some data as follows:

Time Cp observation Pd observation

0 0 -

3 - .5

5 10 -

7 - .7

10 15 -

There will be 2 data records in the control stream, for the PD observation at time 3. The records will start like this:

Time DV SLO INT

3 .5 2 0

7 .7 1 5

Because the line interpolating the Cp from time 0 to 10 has int=0, and slope =2, and the line interpolating the Cp from time=5 to 1ime=10 has intercept 5 and slope=1.

The approriate time scale is the actual time scale (as it is in the time column, above). DES advances the solution of the differential equations using a variable called T. Thus, in $DES, if the "effect compartment" is A(1), you need have only the code

DADT(1) = KEO*(INT+SLO*T - A(1))

to "convolve" the "connect the dots" interpolating function with the monoexponential of the effect comp[artment. Then in $ERROR, if A(1) is the default observation compartment, you write

Y = PD_model(F) + error_model

The DV are all PD observations, of course, and the step that computes the slopes and intercepts for each record can be done just once, before the nonmem run as these never change (the PK data is treated as fixed).

Things get a little trickier (but not much) if the "0" time has a Cp (e.g., steady-state), since then you must initialize A(1) to the approriate (e.g., steady-state) Ce.

===========================================

****

**From alison Thu Sep 26 13:41:09 1996
**

Ruedi Port gives this example of a recursive $PRED:

$PRED

IF (TIME.EQ.0) Celast = 0 ; or some other initial value

Cenew = ... Celast ... (function of Celast and the PK parameters)

effect = ... (function of Cenew and more parameters

Y = ... effect ... EPS( )

Celast = Cenew

This can only work with individual data, i.e., when there are no inter-individual (population) etas.

But Rik's code was for population data.

Maybe someday NM-TRAN will be able to handle a recursive $PRED, but this is far off (not even with NONMEM V).

It is necessary to write one's own PRED for the recursive case with population data. If the functions are relatively simple and the dosing is simple, this is not so very difficult a task. Information in the User's Guides (I and VI) would be helpful.

Alison Boeckmann

****

**From n.holford@auckland.ac.nz Thu Sep 26 14:17:46 1996
**

>

> It *is* possible to use plines in $DES.

> I have some coding for this as part of my lecture notes

> in my Advanced PK course, and if I find the time, I will

> try to make this available.

We look forward to this...

>

> But, this is not necessary, as I wrote to Sam yesterday.

> A linear spline is adequate to represent Cp.

A linear spline *may* be adequate to represent Cp. I personally prefer the idea of a spline with some curvature to it so that I can have a guess at the peak Cp that might be higher than anything I actually measured. The linear interpolation is fine if you have plenty of points. What seemed to be attractive about using CLS is that it was curvy and had other constraints appropriate for describing a pharmacokinetic function which would help when Cp data was relatively sparse. When info is sparse I think one should use any prior info to assist. In this case I am pretty sure that a curvy spline is going to be better than a linear spline.

--

Nick Holford, Dept Pharmacology & Clinical Pharmacology

University of Auckland, Private Bag 92019, Auckland, New Zealand

email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556

http://www.phm.auckland.ac.nz/Staff/NHolford/nholford.html

****

**From rs@chdr.leidenuniv.nl Thu Sep 26 23:53:39 1996
**

Dear Ken, you wrote

> Rik,

>

> You might want to consider the semicompartmental modeling approach that I have

> been developing. The semicompartmental approach is a solution to Sheiner's

> effect-site link model based on a noncompartmental model for Cp (piecewise

> linear or log-linear model) and is easily implemented in standard nonlinear

> regression packages such as NONLIN, NONMEM, and SAS NLIN. I was motivated to

> develop this approach precisely for the reason you indicated in your message,

> ie., when the kinetic profile is not accurately described by standard

> compartmental pharmacokinetic models. The catch with my approach is that you

> need to have enough times points at properly spaced intervals such that the AUC

> can be accurately estimated by linear and/or log-linear trapezoidal rule

> calculations. Here is the reference for this approach:

>

> Kowalski, K.G. and Karim, A. A Semicompartmental Modeling Approach for

> Pharmacodynamic Data Assessment. J. Pharmacokin. & Biopharm., 23:307-322

> (1995).

>

> If you decide to try the semicompartmental modeling approach, let me know if I

> can be of further assistance.

>

> Ken Kowalski

> G.D. Searle

> Skokie, IL

>

>

But after my request for a recursive implementation (which from your paper I gather is necessary) Alison wrote as a response to Ruedi's reply:

>Ruedi Port gives this example of a recursive $PRED:

>

>$PRED

>

> IF (TIME.EQ.0) Celast = 0 ; or some other initial value

> Cenew = ... Celast ... (function of Celast and the PK parameters)

> effect = ... (function of Cenew and more parameters

> Y = ... effect ... EPS( )

> Celast = Cenew

>

>This can only work with individual data, i.e., when there are no

>inter-individual (population) etas.

>

>But Rik's code was for population data.

>

>Maybe someday NM-TRAN will be able to handle a recursive $PRED,

>but this is far off (not even with NONMEM V).

>

>It is necessary to write one's own PRED for the recursive

>case with population data. If the functions are relatively simple

>and the dosing is simple, this is not so very difficult a task.

>Information in the User's Guides (I and VI) would be helpful.

>

>Alison Boeckmann

>

Does this mean that the only way to implement it in NONMEM is by using a differential equation in connection with slopes and intercepts for the interpolating straight lines? Or can you write

down an explicit solution after all, since you say that it is easily implemented in NONMEM?

Kindest regards,

Rik Schoemaker,

CHDR, Leiden, NL

****

**From R.Port@dkfz-heidelberg.de Fri Sep 27 03:35:30 1996
**

Rik Schoemaker's message of Sept 26

My message of Sept 26 including this idea for a recursive $PRED

> IF (TIME.EQ.0) Celast = 0 ; or some other initial value

> Cenew = ... Celast ... (function of Celast and the PK parameters)

> effect = ... (function of Cenew and more parameters

> Y = ... effect ... EPS( )

> Celast = Cenew

Alison's message of Sept 26

Hi Alison,

Thanks to your comment I see that the above code won't work with population data, at least when the FO method is used. I'm wondering whether it could be applicable with the FOCE method. What is your answer?

Thanks! Ruedi

-------------------------------------------------------------------------------

R.E. Port, Dept. 0420, German Cancer Research Center

P.O. Box 10 19 49, D-69009 Heidelberg

phone: x49-6221 42-3385

-3347

fax: -3346

e-mail: r.port@dkfz-heidelberg.de

****

**From R.Port@dkfz-heidelberg.de Fri Sep 27 03:47:36 1996
**

Rik Schoemaker's message of Sept 26

My message of Sept 27 including this idea for a recursive $PRED

> IF (TIME.EQ.0) Celast = 0 ; or some other initial value

> Cenew = ... Celast ... (function of Celast and the PK parameters)

> effect = ... (function of Cenew and more parameters

> Y = ... effect ... EPS( )

> Celast = Cenew

Alison's message of Sept 26

My message of Sept 27

> Hi Alison,

>

> Thanks to your comment I see that the above code won't work with population

> data, at least when the FO method is used. I'm wondering whether it could be

> applicable with the FOCE method. What is your answer?

Hi Alison,

of course, for FOCE, I was thinking of a code that includes eta's:

IF (TIME.EQ.0) Celast = 0 ; or some other initial value

Cenew = ... Celast ... (function of Celast,

the mean population PK parameters, and some eta's)

effect = ... (function of Cenew, more parameters (more eta's))

Y = ... effect ... EPS( )

Celast = Cenew

What do you think?

Thanks! Ruedi

-------------------------------------------------------------------------------

R.E. Port, Dept. 0420, German Cancer Research Center

P.O. Box 10 19 49, D-69009 Heidelberg

phone: x49-6221 42-3385

-3347

fax: -3346

e-mail: r.port@dkfz-heidelberg.de

****

**From alison Fri Sep 27 10:30:50 1996
**

Kyungsoo Park, Rik Schoemaker, Ken Kowalski, Reudi Port, Lew Sheiner, and others have been discussing various complicated topics in modelling. I do not want to comment on these topics, because I am not familiar enough with them. The only thing I am comfortable commenting on is this (most recent) code and question from Reudi Port:

> I see that the above code won't work with population

> data, at least when the FO method is used. I'm wondering whether it could be

> applicable with the FOCE method.

> of course, for FOCE, I was thinking of a code that includes eta's:

>

> IF (TIME.EQ.0) Celast = 0 ; or some other initial value

> Cenew = ... Celast ... (function of Celast,

> the mean population PK parameters, and some eta's)

> effect = ... (function of Cenew, more parameters (more eta's))

> Y = ... effect ... EPS( )

> Celast = Cenew

This cannot be implemented correctly in abbreviated code. Unfortunately, NM-TRAN (in NONMEM IV) does not recognize that there is something wrong.

With population data and both FO and FOCE, the objective function will be computed incorrectly. With FOCE, I think it will be even worse, and that the conditional estimation will also be done incorrectly.

With individual data and FO, the computations will be correct, because there will be no etas in the computation of Cenew and Celast. (FOCE is not possible with individual data).

As to what this implies for the complicated models that have been discussed, I cannot say in general.

Here are some details that *might* make this more clear. Skip them if they confuse anyone.

At present, NM-TRAN makes a single pass thru a given block of abbreviated code. When it sees "Cenew = ... Celast", it does not know that Celast will become dependent on etas later in the block, and does not compute the partial derivatives correctly.

This is true of $PRED, $DES, $PK, and $ERROR blocks. With the latter three, of course, we are using PREDPP, and PREDPP can handle a recursive solution internally; compartment amounts *and their derivatives* are carried forward and updated correctly from record to record. But the PK parameters (and the quantities used in differential equations and in the $ERROR block) must be coded in a non-recursive way.

That is, any left hand quantity in abbreviated code that depends on random variables (ETA and/or EPS) must be computed EXACTLY ONCE AND UNCONDITIONALLY.

If an either/or choice is involved, indicator variables (e.g., Q and 1-Q in many examples) can be used, because this rule is obeyed.

But if the computation invoves random variables computed with a prior data or event record, it cannot be done correctly.

****

**From KATYAG@otsuka.oapi.com Tue Oct 1 16:14:16 1996
**

Kyungsoo and others who have the experience with CLS program,

can the program be used with more complex data than in the examples provided with the program? Specifically:

1. What if plasma concentrations spread all over the dosing interval, but for each particular individual observations are concentrated in a small range, say, one to two hours. May I expect reasonable estimates of AUC? If yes, how many break points to use, where ?

2. What if there are no observations at 0 or/and at the end of a dosing interval for steady state. Should I still have break points there? Is there another way to constrain the spline at these points?

3. If there is a positive answer to the previous question, then another complication: can the steady state requirement at times 0 and T be combined with the requirement of a decreasing tail at times >T ? I.e., if being at a steady state a patient skips the last dose and has a plasma measurement later. Is there a way to handle it?

Thanks in advance.

Katya Gibiansky

Otsuka America Pharmaceuticals, Inc.

2440 Research Blvd, Rockville,

MD 20850

USA

E-mail: katyag@otsuka.oapi.com

Phone: (301)-527-4911

Fax: (301)-212-8582

****

**From sambol@itsa.ucsf.edu Tue Oct 1 17:51:32 1996
**

Dear Rik,

Some more comments regarding kinetic interpolation and effect compartments.

We ran into the same problem (erratic conc.-time profile, but need to obtain predicted "effect site" concentrations to feed into PD model). We developed an empiric convolution method to obtain predictions of "effect site" concentration at time points corresponding to PD measurements. The method was used in the context of a population analysis (using NONMEM) of concentration vs. pain relief score of an analgesic. (Liu & Sambol, Pharm. Res., 12: 438-445, 1995; see also Liu & Sambol, Bulletin of the International Statistical Institute, 50: 716-717, 1995) Our goal is essentially the same as that which Lew described, but the implementation involves an explicit solution.

The procedure involves a piecewise linear approximate to the concentration-time profile and convolution, carried out exactly, based on this approximate. The feature of this implemenation we like in particular is that it can be performed within $PRED. The general expression for this equation is provided in the paper cited above.

An example of the relevant NONMEM code with time points 0=T0<T1<T2<T3 and observed plasma concentrations C1 and C3 (without C2, and with the desire to predict Ce at T2), is as follows:

------------

$PRED

...

A1=0

B1=C1/T1

IF (TIME.LE.T1) THEN

CE=KEO*(B1*(KEO*T1-1.0+EXP(-KEO*T1))/(KEO**2))

ENDIF

IF(TIME.LE.T3.AND.TIME.GT.T1) THEN

B2=(C3-C1)/(T3-T1)

A2=C3-B2*T3

CE=B1*(KEO*TIME-1.0+EXP(-KEO*TIME))/(KEO**2)

CE=CE+(A2-A1)*(1.0-EXP(-KEO*(TIME-T1)))/KEO

CE=KEO*(CE+(B2-B1)*(KEO*T-1.0-(KEO*T1-1.0)*EXP(-KEO*(TIME-T1)))/(KEO**2))

ENDIF

...

-------------

Note: Cn and Tn are data fields (read into NONMEM with either $DATA or verbatim code), although you don't need records for Tns and Cns where PK data are missing (i.e., measurements like C2 in the example). You also need a TIME field in $DATA with values corresponding to effect measurements. You can handle the fact that different individuals may have different numbers of concentration-time data points using "IF" statements. Finally, as you can imagine, if the distance between concentration observations is large (at certain locations), this method may not work so well.

Chui Y Liu and Nancy Sambol

Nancy C. Sambol, Pharm.D.

Department of Biopharmaceutical Sciences

University of California San Francisco

San Francisco, CA 94143-0446

phone: 415-476-8884

fax: 415-476-9330

****

From kyungsoo Wed Oct 2 12:06:30 1996

Subject: Re: CLS Program

1. From the population point of view, sampling spreading over time could be better than sampling at the same times for all individuals. So you would probably get some estimate of AUC. And in that case, one way of chooing the breakpoints would be locating them at the quantiles of sampling times. Regarding the number of breakpoints, since the study is not well desigened, you might want to use some model selection procedure (e.g., Akaike criterion) using the pooled data (no interindividual error; ETA = 0).

2. Yes, you need breakpoints there.

3. The code in the repository is already constrained to achieve a decreasing tail at times >T for steady state (see constraint 2 in readme.txt)

Hope this helps.

Kyungsoo Park

****

**From kyungsoo Wed Oct 2 13:03:54 1996
**

Regarding what Katya Gibiansky wrote:

> 3. If there is a positive answer to the previous question, then another

> complication: can the steady state requirement at times 0 and T be

> combined with the requirement of a decreasing tail at times >T ? I.e., if

> being at a steady state a patient skips the last dose and has a plasma

> measurement later. Is there a way to handle it?

I was trying to say

3. The code in the repository is constrained to achieve a decreasing tail at time = T for steady state (see constraint 2 in readme.txt) and assumes that a patient does NOT skip the last dose.

If he does, one may need to (exponentially) extrapolate CLS beyond the end of a dosing interval to get the prediction for the later measurement ...

Kyungsoo Park

****

**From kyungsoo Wed Oct 16 12:23:59 1996
**

Dear NONMEM users:

Since the CLS program was released last month, several NONMEM users have raised a question that if it can also be used in $DES for PK/PD modeling, as well as in $PRED for PK modeling.

The program has been REVISED in that way. That is, the CLS subroutines can now be called in $DES to obtain the individual's plasma concentration predictions at the effect measurement times. In more details, the new version can be used for:

1. PK modeling

1.1. Single-dose

1.2. Steady-state

2. PK/PD modeling

2.1. Single-dose: Effect is indirectly linked to plasma concentration. (e.g., effect compartment)

2.2. Single-dose: Effect is directly linked to plasma concentration.

2.3. Steady-state: Effect is indirectly linked to plasma concentration.

The files for the new version are available in NONMEM.DIR/CLS.DIR (same directory as before) by anonymous ftp from the NONMEM repository or from the NONMEM users network:

ftp pkpd.icon.palo-alto.med.va.gov

http://pkpd.icon.palo-alto.med.va.gov

Thanks.

Kyungsoo Park, Davide Verotta, and Lewis B. Sheiner

University of California, San Francisco

Room C255, Box 0626

San Francisco, CA 94143

E-mail: kyungsoo@c255.ucsf.edu