Re: [NMusers] Steady-state (SS) and $DES veeeery slow

From: Alison Boeckmann <alisonboeckmann_at_fastmail.fm>
Date: Thu, 19 May 2016 07:23:12 -0700

Dear Paolo,
 
As Bob explained, ZSPOW uses the  TOL defined by the user in the
$SUBROUTINES record to determine precision.  ADVAN9 and ADVAN13 also u=
se
a separate absolute tolerance ATOL, which is by default set to 12, but
you can modify it with $EST … ATOL= , and see if it affects the
precision of the SS evaluation.
You could also try using  ADVAN6/SS6.
 
I can't understand why the SS option is taking so long with
ADVAN13/SS13.  The table of results you report do not correspond to the
way SS is computed in PREDPP, and I can't understand where you got the
numbers in the table.
 
You say "the amount in the absorption compartment AA1 is literally
numerical noise after II hours". But PREDPP never reports this amount,
because SS amounts are reported *after* a dose has been administered.
 
The way SS with multiple dosing is computed is described in the May 2015
version of Guide VI PREDPP.  This is supplied with the latest alpha
versions of NONMEM7.4.  See Chapter V. Event Records, Section F.4
Implementation of Steady State.  There are three calls to ADVAN with
TIME=0 to obtain the initial estimate.  After that, the root finder
ZSPOW makes as many calls as it needs, but they are not necessarily
sequential in TIME.

Can you please send me your control stream and a fragment of the data so
I can figure out what is going on?
 
Alison
 
PS. In case you don't have the May 2015 version of guide VI PREDPP, here
    is an extract:
(The guide still needs to be updated for ADVAN13, but the same
process is used.)
 
=========================
When the model is defined by differential equations (SS6, SS8, SS9),
PREDPP cannot use an ana- lytic solution, even if one happens to exist.
Instead, the solutions are found by a numerical technique using ZSPOW, a
root-finding subroutine from IMSL.
With multiple dosing, ZSPOW searches for a vector of compartment amounts
and their eta deriva- tives such that the initial and final (end of
dosing interval) vector is the same. ADVAN6, ADVAN8, and ADVAN9 perform
a numerical integration of the differential equations (and in the case
of ADVAN9, algebraic equations). The ADVAN routine is used to advance
(integrate) the state vector A from time 0 to time II with the
appropriate dose pattern. Let AII represent the state vector after
advance to time II with the appropriate dose. A search is made for Ass
such that Ass - AII = 0
If SS=3, the state vector A contains user-supplied initial estimates.
Otherwise, the SS routines sets A=0 and makes three calls to ADVAN. Each
call advances the state vector from t=0 to t=II with the appropriate
dosing pattern. The result is the initial estimate.
Once the SS solution is found, the SS routine adds the final bolus dose
to the state vector, or starts the final infusion.
===================
On Wed, May 18, 2016, at 03:15 AM, Paolo Denti wrote:
> Dear NMUsers (and developers),
> I am trying to speed up our run times when using SS in ADVAN13 models
> (user defined differential equations), and I would like to share some
> thoughts to get some feedback.
>
> We work on steady-state PK data, so ideally we would like to use the
> SS option, but the run times whenever we use ADVAN13 become
> unfeasibly long (even 20 times more), and this even for models that
> should reach steady state in 2-3 doses.
> As an alternative we end up using ad-hoc "patch-up" solutions, like
> initialising compartments, or just adding 4-5 doses "manually" in the
> dataset, but this is a bit tedious/tricky.
> I am writing to see if there is a way to speed up the SS feature.
>
> I decided to look into what was going on by asking NONMEM to simulate
> a SS PK profile and print out all the temporary iterations.
> It seems like NONMEM first opens a separate "SS session" used to work
> out the SS amounts in each compartment. The model is taken to SS by
> repeatedly giving doses every II, until the system is deemed to have
> reached steady-state. At this point, NONMEM goes back to the "main
> session", initialises all compartments to the amounts found in the
> "SS session", gives the final dose, and moves on with the analysis.
>
> Nothing surprising here, and it's understandable that things take
> longer with SS, cuz the calculation of these SS amounts implies
> solving the differential equations for all the extra time of the "SS
> session". What I found odd though, is that the number of doses that
> NONMEM uses to reach steady-state seems to me much higher than
> needed. In my test I used a simple 1-cmpt KA model coded in ADVAN13
> and a drug with a half-life of 3.5 hours, so I was expecting 2 dosing
> intervals (48 hours) to be more than enough to get to steady state,
> as maths says the amounts should be 99.993% there. NONMEM instead
> used 13 doses in the example below.
>
> The amounts are reported in the table below. After iteration 3 the
> amount in AA2 just changes by tiny values, arguably comparable to
> numerical noise.
>
> #DOSE    T    TOT_TIME    A=
A2
> 0    0    0
> 1    0    24    4.114874
> 2    0    48    4.148738
> 3    0    72    4.149017
> 4    0    96    4.149017
> 5    0    120    4.149017
> 6    0    144    4.149019
> 7    0    168    4.149019
> 8    0    192    4.149019
> 9    0    216    4.149019
> 10    0    240    4.149019
> 11    0    264    4.149019
> 12    0    288    4.149019
> 13    0    312    4.149019
>
> Does anyone know what stopping criterion NONMEM uses to call it SS
> and move on? Is there a way to relax it? I think 0.1% would be fine
> in most practical cases - at least for preliminary runs - and in this
> example it would save 80% of the run time. Plus if one keeps in mind
> that this is a numerical solver, it is not clear how "real" the wee
> digits are, obviously depending on TOL.
>
> The other tricky thing I found is that the amount in the absorption
> compartment AA1 is literally numerical noise after II hours, meaning
> that for some entries it is even negative (e.g. -2.08038E-15) and
> jumps between positive and negative. Could it be that this is what is
> tricking NONMEM's stopping criterion to detect SS? Maybe if the
> stopping criterion only asks for a relative change <TOL, it will
> struggle to achieve that with values that change sign.
>
> Any settings that may help speed things up? Something like tinkering
> with TOL and ATOL?
> Or maybe is it possible to set a maximum to the number of doses that
> NONMEM tests to reach steady-state?
>
> Sorry for the nerdy topic, and thanks for any advice!
> Ciao,
> Paolo
>
> PS And big thanks to our PhD student Maxwell who ran all the tedious
> simulations!
>
>
> -- ------------------------------------------------ Paolo Denti, PhD
> Pharmacometrics Group Division of Clinical Pharmacology Department of
> Medicine University of Cape Town K45 Old Main Building Groote Schuur
> Hospital Observatory, Cape Town 7925 South Africa phone: +27 21 404
> 7719 fax: +27 21 448 1989 email: paolo.denti_at_uct.ac.za ------------------=
------------------------------
> Disclaimer - University of Cape Town This e-mail is subject to UCT
> policies and e-mail disclaimer published on our website at
> http://www.uct.ac.za/about/policies/emaildisclaimer/ or obtainable
> from +27 21 650 9111. If this e-mail is not related to the business of
> UCT, it is sent by the sender in an individual capacity. Please report
> security incidents or abuse via csirt_at_uct.ac.za
 
--
Alison Boeckmann
alisonboeckmann_at_fastmail.fm
 
 


Received on Thu May 19 2016 - 10:23:12 EDT

This archive was generated by hypermail 2.3.0 : Fri Sep 27 2019 - 16:50:35 EDT