**From: rs@chdr.leidenuniv.nl
**

Dear fellow users,

I'm trying to model concentration effect relationships for propofol and midazolam and I run into proteresis. I have a more or less linearly increasing concentration using a target controlled infusion (TCI) regime (with a duration of somewhere between 30 and 90 min) followed by a 30 min plateau phase after which the infusion terminates. The TCI means that I cannot do ordinary PK modeling because of the continuously varying input rate and I hoped to be able to model using ordinary linear interpolation, specifying slope and intercept of the interpolating segments. This works very well for an ordinary effect compartment situation with hysteresis. However, I have venous concentrations and (especially for propofol) proteresis which I assume is due to arterio-venous differences.

The problem is (I assume) that for hysteresis the measured concentrations drive the hypothetical effect compartment concentrations that are related to the effect while in this case the arterial concentrations that drive the venous concentrations are related to the effect.

Do you agree about the assumption of arterio-venous differences and does anyone have suggestions for modeling this specific situation?

Thanks in advance,

Rik Schoemaker

CHDR, Leiden, The Netherlands

******
**

**From: Mark Sale <msale@medlib.iaims.georgetown.edu>
**

Rik,

You're correct that in the explicit ADVANS it isn't easy to do a continuosly chaning infusion rate. However, it can be done in the differential equation advans, by making an infusion rate a function of time, and checking to see if all drug has been infused. This could then explain the arterial compartment, and the effect compartment and venous compartment could be attached to that with (linear?) rate constants. At the end of this is a control stream and data file. But I'd geuss other may have more clever ways of doing this.

As far as you suggestion that proteresis can result from A-V difference, where the effect compartment is the arterial compartment and the sample is collected from the venous compartment, you're correct, in a rapidly changing system (i.e., rate constants on the order of 2/min) this can occur, even in the absence of tissue extraction. The other common explaination is tolerance. I'd refer you to the "classic" paper descriibing tolerance to nicotine by Porchet and Sheiner.

Mark

Mark Sale M.D.

Assistant Professor of Medicine

Georgetown University

Washington DC 20007

202-687-8242

msale@medlib.georgetown.edu

$PROB infusions

$INPUT TIME AMT DV

$SUBROUTINES ADVAN6 TOL = 5

$DATA DATA

$MODEL

COMP = (DEPOT,DEFDOSE)

COMP = (CENTRAL,DEFOBS)

$PK

RATE = 1 ;THIS DEFINES THE RATE, COULD ALSO BE INCLUDED IN DATA SET

S2 = THETA(1)

KEL = THETA(2)

$DES

;FIRST CHECK TO SEE IF ANY DRUG LEFT IN COMPARTMENT

;THE TOTAL DOSE INFUSED IS LISTED IN THE DATA SET

;THIS IS THEN INFUSED, AT A VARYING RATE UNTIL IT

;IS GONE

IF(A(1).LE.0) THEN

IND = 0

ELSE

IND =1

END IF

DADT(1) = -RATE*IND*T ;INFUSION RATE IS FUNCTION OF TIME

DADT(2) = RATE*IND*T-KEL*A(2)

$ERROR

Y = F+ERR(1)

$THETA

(0,1) ;VOLUME

(0,1) ;KEL

$OMEGA 1

$ESTIMATION MAX=9999 PRINT=2

$TABLE TIME IND NOPRINT FILE = OUT.XLS

.......................

data

0 50 .

0.5 . 3

1 . 7

1.5 . 10

2 . 12

2.5 . 17

3 . 20

3.5 . 22

4 . 25

5 . 30

6 . 35

7 . 40

8 . 45

9 . 50

10 . 55

11 . 48

12 . 44

13 . 40

******
**

**From: lewis@c255.ucsf.EDU (LSheiner)
**

Rik -

A further remark on the AV difference and proteresis:

The paper in which we worked out a semi-parametric approach to this problem is

Verotta, Beal, & Sheiner, Am J Physiol 256 (Reg Integ Comp Physiol 25) R1005-R1010, 1989.

The idea, as you correcly surmise, is to model an unobserved arterial compartment which drives both the venous compartment and the effect compartments, and in your case the venous compartment will be "shalower" than the effect compartment.

We put forward a method that is semiparametric and involves (constrained) deconvolution ... this is not going to be easy (possible?) using NONMEM. So, a more parametric version of our model might proceed as follows:

DADT(1) = KVO*A(3) - KVO*A(1) ; Venous Conc, Cv

DADT(2) = KEO*A(3) - KEO*A(2) ; Ce

DADT(3) = K43*A(4) - (K30+K34)*A(3) ; Arterial Conc, Ca

DADT(4) = -K43*A(4) + K34*A(3) ; "Tissue" Conc

The above models Ca as biexponential. Note that it has no loss from A(3) to A(1) or A(2) - they are both "hypothetical" even tho Cv is observed (that is, A(4) is supposed to take care of the kinetics being non-first-order)

For identifiability, S1 should be fixed to 1 (we observe Cv), and S4 should be estimated (the model requires SS Cv = SS Ce = SS Ca).

Note:

1) In principle, you have to fit the PK and PD simultaneously - you can't do the PK first and fix on it since it takes the PD proteresis to define the Ca kinetics. HOWEVER, you can ASSUME a value for KVO from knowledge of the perfusion of the site from which the venous drainage is sampled. It doesn't matter very much if this is right or not, since all it has to be is sufficienlty "slow" so that the estimated arterial kinetics "get ahead" of the

effect (i.e., they have to be such that estimated Ca vs effect exhibits no proteresis; it can, of course, exhibit hysteresis, which will be taken care of in the PD fit by KEO.

2) If you fit simultaneoulsy, it is likely that KEO -> infinity, in which case, as you suggested, Ce cannot be distinguished from Ca, and the model requires one less parameter. It will be unidentifiable as I have written it. Therefore, if yotu want to do a simultaneous fit, you might start with fixing KEO to a very large value, and then, once you have a fit, seeing if you can let it go free.

Good luck.

******
**

**From: alison@c255.ucsf.EDU (ABoeckmann)
**

Most of you have seen Rik Schoemaker's question, and responses from Mark Sale and Lew Sheiner.

I'm interested in the PK part of the model.

Lewis remarked to me that the PK part of the model is not of great interest and can be handled with a series of constant rate infusions. During the time of rapid rise (say 30 mins), use several short infusion records (6 or so) with fixed rate infusions having RATE = the mean infusion rate during that 5 minute period, and the known AMT, followed by a longer (plateau) infusion, again with the known RATE and AMT. This implements the infusion via a sort of step function.

ADVAN1 can be used, or, if Differential eqns. are needed for the rest of the model, use the constant rate infusions into the central comparment of the model.

However, I am interested in Mark's attempt to model variable rate infusion with differential eqns.

His D.E.'s have a flaw that we have not perhaps warned people about in Guide VI. Within an integration interval, the D.E.'s must be continuous and smooth. His D.E.'s have a discontinuity when A(1) becomes 0. They may seem to work, but this is not guaranteed.

Discontinuities in D.E.'s should occur only at event times (or at non-event dose times such as lagged doses).

He needs event records at the start and end of the plateau. This can easliy by done because the times (T1 and T2) are known to the experimenter.

I think it is easiest to introduce a new Covariate PLAT with 3 values: 0 (before plateau), 1 (during plateau), 2 (after plateau)

What we want is this:

RATE=T from 0 to T1 ; fast rise

RATE=T1 from T1 to T2 ; plateau rate

RATE=0 from T2 onwards ; normal elimination

Thus:

IF (PLAT.EQ.0) RATE=T

IF (PLAT.EQ.1) RATE=T1

IF (PLAT.EQ.2) RATE=0

T1 is the time when the target concentration was reached.

T2 is the time when the infusion ends.

Suppose T1=30 and T2=60.

Use PLAT=0 in event records up to and including 30, and PLAT=1 in the event records subsequent to 30, so that PLAT=0 governs the advance up to time 30, and PLAT=1 governs the advance past time 30. Similarly, PLAT=1 at TIME 60, PLAT=2 at larger times.

The AMT of drug for CMT 1 is effectively infinite - it simply acts as a resevoir of drug.

$INPUT ... TIME PLAT AMT ...

0 0 999999

30 0 .

30 1 . "documentation" record

60 1 .

60 2 . "documentation" record

The "docuementation" records have no effect on the predictions because there is no advance when TIME does not change. However, they do serve to remind people who look at the data that this is the actual time that the covariate PLAT changes.

Control stream and DATAPL follow. They are based on Mark Sale's control stream and data file. IND1 changes after time T1: it switches beteen T and T1 as the rate of infusion. IND2 changes after time T2. It turns off the infusion altogether.

I've appended extra other-type event records at later points in time so that one can see the compartment amounts beyond the last observation, to verify that the plateau exists.

I've also put some "debugging" output RR, RATE, etc. in the table.

There is no $ESTIM record - no point in Estimation till we are sure the model is doing the right thing!!!

In Rik's case, T1 and T2 are definitely known. It would be possible to estimate T1 and T2 using NONMEM. However, this is more complicated. Lewis says that people who need continuously varying input rates know what the rates and times are and don't need to estimate them, and there's no point solving a problem that doesnt exist, so I'll leave it at that.

$PROB infusions

$INPUT TIME AMT DV PLAT EVID

$SUBROUTINES ADVAN6 TOL = 5 LIB

$DATA DATAPL

$MODEL

COMP = (DEPOT,DEFDOSE)

COMP = (CENTRAL,DEFOBS)

$PK

S2 = THETA(1)

KEL = THETA(2)

IND1=0

IND2=1

IF (PLAT.GE.1) IND1=1

IF (PLAT.EQ.2) IND2=0

T1=30

$DES

RR=T*(1-IND1)+T1*IND1

RATE=IND2*RR

DADT(1) = -RATE

DADT(2) = RATE - KEL*A(2)

$ERROR

Y = F+ERR(1)

A1=A(1)

A2=A(2)

$THETA

(0,1) ;VOLUME

(0,1) ;KEL

$OMEGA 1

;$ESTIMATION MAX=9999 PRINT=2

$TABLE TIME IND1 IND2 RR RATE FILE = OUTP.XLS

$SCAT PRED VS TIME

$SCAT (A1,A2) VS TIME

============= DATAPL ===============

0 9999999 . 0 1

0.5 . 3 0 0

1 . 7 0 0

1.5 . 10 0 0

2 . 12 0 0

2.5 . 17 0 0

3 . 20 0 0

3.5 . 22 0 0

4 . 25 0 0

5 . 30 0 0

6 . 35 0 0

7 . 40 0 0

8 . 45 0 0

9 . 50 0 0

10 . 55 0 0

11 . 48 0 0

12 . 44 0 0

13 . 40 0 0

20 . . 0 2

30 . . 0 2

30 . . 1 2

35 . . 1 2

40 . . 1 2

50 . . 1 2

60 . . 1 2

60 . . 2 2

61 . . 2 2

65 . . 2 2

70 . . 2 2

89 . . 2 2

******
**

**From: Peter Wright <p.m.c.wright@ncl.ac.uk>
**

Rik and other NMusers

>The TCI means that I cannot do ordinary PK modeling because of the

>continuously varying input rate and I hoped to be able to model using

>ordinary linear interpolation, specifying slope and intercept of the

>interpolating segments. This works very well for an ordinary effect

>compartment situation with hysteresis. However, I have venous

>concentrations and (especially for propofol) proteresis which I assume

>is due to arterio-venous differences.

>

>The problem is (I assume) that for hysteresis the measured concentrations

>drive the hypothetical effect compartment concentrations that are related

>to the effect while in this case the arterial concentrations that drive the

>venous concentrations are related to the effect.

>

>Do you agree about the assumption of arterio-venous differences and

>does anyone have suggestions for modeling this specific situation?

>

I have faced these problems and dealt with them in the following ways.

After an experiment conducted using target controlled infusions I created data files with multiple dosing events reconstructed from the output of Stanpump. This is rather crude and run times were slow as might be expected. However it worked. I believe there is a utility on the Stanford PK/PD server which can create a list of dose events from a Stanpump output in a relatively easy way (though I have not myself seen it).

A fairly simple way to model arterio-venous differences in a quasi compartmental fashion is to create an effect type compartment (which follows arterial concentration via a rate constant). I call this the gradient compartment. Then model venous concentrations as being contributed to by arterial concentrations and concentrations in the gradient compartment such that the relative proportions add up to 1. This models venous concentrations as initially less then arterial then later greater than arterial, but with AUC identical. There are two parameters in the model. In the situation of peripheral elimination (e.g. nitroglygerin,atracurium) an additional parameter will be required to adjust the " total proportion" down from 1. In the situation desribed above this will only work if PK and PD are modeled simultaneously because you have data only from the arterial side. The code to achieve this model when data are available for from both artery and vein is as follows.

Arterial concentrations are compartment 1. The "gradient" compartment is 4. Since you have concentrations only from the venous side you will be able to simplify this somewhat.

$ERROR Y1 = F * (1+ERR(1))

Cp = A(1)/V1

Cg = A(4)/V4

PROPG = THETA(8)

PROPA = (1 - THETA(8))

Cv = (Cg * PROPG) + (Cp * PROPA)

Y2 = Cv * (1+ERR(2))

Q1 = 0

Q2 = 0

IF (CMT.EQ.1) Q1 = 1

IF (CMT.EQ.4) Q2 = 1

Y = Q1*Y1 + Q2*Y2

If there are any mistakes in this please let me know.

Peter Wright

University of Newcastle upon Tyne