Jump to content


Photo

Another comparison of results NLME vs NOMEM


  • Please log in to reply
6 replies to this topic

#1 jjohnl

jjohnl

    Member

  • Members
  • PipPip
  • 10 posts

Posted 15 May 2019 - 02:22 AM

Hi,

 

I received a report which used NOMEM to generate a Pop PK model from a study containing 8 subject plasma concentration dataset of 300 mg Oral tablet (no covariance was used).   NOMEM successfully generated a 1st order 2 compartment no lag time Pop PK model.  FOCE method was employed for all non-linear mixed effect modeling. The inter-individual random effects on the parameters were modeled assuming the multiplicative form (0.1). The residual variability was modeled according to a proportional error model.

 

The POP Pk parameter (RSE%) generated from NONMEM were:

ka=3.22 (19.3), V = 3880 (11.9), V2 = 3170 (8.71), CL = 324 (7.13), Q = 485 (18.4)

 

I tried Phoenix NLME and ran POP PK (1st order 2 compartment no lag) using each algorithm.  All algorithm except FOCE L B gave much smaller Ka values (Ka ~ 1.x) than that of NOMEM; The other 4 parameters were more or less in agreement with NOMEM's.

The Ka from FOCE L B was ~ 3.7

 

Can someone let me know if I carried out the POP Pk fitting correctly?

Please see attached data (one of the Pop PK fit I even used NOMEM parameters are initial estimates just for fun...)

 

Thanks

J

 

Attached Files


  • Stevensi, Thomasgaks, Keithked and 2 others like this

#2 smouksassi1

smouksassi1

    Advanced Member

  • Members
  • PipPipPip
  • 231 posts
  • LocationMontreal

Posted 15 May 2019 - 06:43 AM

Hi J,

 

In the project I did not see NONMEM control file and output to be sure we are using the same dataset and same options.

NONMEM has also lot of fitting algorithm and options. FOCE-LB is known to be problematic with non additive errors so unless you are looking for a quick fit to start avoid it and stip to FOCE-ELS and QRPEM.

 

N =8 is pretty small to estimate variances the model can be quite unstable and overparametirzed.

 

Even when you have tons of data a two compartment absorption model can be hard to show with parameters that are identifiable there is multiple solution and some parameters might flip flop within individuals unless you code the model in a way to avoid it

and what is true for one individual might be different when you fit a pop model ( some parameters become identifiable given some random effects structures)

https://hal.archives...251986/document

 

 

( you can always tweak nonmem to give results similar to nlme or sas nlmixed or tweak nlme to give nonmem results , this does not mean that one is correct and the other no this is the reality of nonlinear fitting you need to check and  decide whether results are trustable and make sense)



#3 Simon Davis

Simon Davis

    Advanced Member

  • Administrators
  • 1,328 posts

Posted 15 May 2019 - 06:58 AM

Hi John, in addtions to Samer's comments I would definitely consider NOT using the MDV flag, it's extravascular data so telling the model that the conc was zero at time zero gives it more information.

ALso I think a Tlag helps a lot with the fit but I'm unsurprised that Ka was problematic, you're essentiallly trying to estimate with just one or two points in the profile.

 

  I would advise using slightly shorter names and focussing more on the structural elements and diagnostic plots etc than the algorithms.  i stuck with FOCE-ELS (but QRPEM is also fine) and tried two models with and without a lag but including the  time zero point.

 

Parameter Estimate CV%

tvKa        4.55         42.35

tvV          4284.59   3.90

tvV2        2732.48   2.80

tvCl         313.06    5.17

tvCl         2413.53  5.32

tvTlag     0.43        4.71

stdev0    0.06         25.51

 

   Simon.

PS I'm not sure why you marked some runs as failed - they have converged although tlag benefits from a lower limit of 0 I think.

Attached Files


Edited by Simon Davis, 15 May 2019 - 07:08 AM.


#4 smouksassi1

smouksassi1

    Advanced Member

  • Members
  • PipPipPip
  • 231 posts
  • LocationMontreal

Posted 15 May 2019 - 08:27 AM

wanted to add that keeping the zero zero points and fitting it is a matter of debate since it is not clear if they were indeed measured and whether they were BLQ and were replaced by zeros also if you have a zero observation your proportional error will not work 



#5 jjohnl

jjohnl

    Member

  • Members
  • PipPip
  • 10 posts

Posted 15 May 2019 - 06:10 PM

Hi Sam,

 

Hi J,

 

In the project I did not see NONMEM control file and output to be sure we are using the same dataset and same options.

NONMEM has also lot of fitting algorithm and options. FOCE-LB is known to be problematic with non additive errors so unless you are looking for a quick fit to start avoid it and stip to FOCE-ELS and QRPEM.

 

 

Attached is the NONMEM ctrl and output (I captured from report) and the NONMEM datafile.  It's pretty straight forward.  EVID --> 0 = plasma samples, 1 = dosing event; MDV --> 0 = data, 1 = (either missing or BQL); CMT was assigned 2 = plasma sample; 1 = dose event. 

 

Thanks

John

Attached Files



#6 jjohnl

jjohnl

    Member

  • Members
  • PipPip
  • 10 posts

Posted 15 May 2019 - 06:22 PM

Hi Simon,

 

PS I'm not sure why you marked some runs as failed - they have converged although tlag benefits from a lower limit of 0 I think.

 

Actually I forgot to ask that question..  For the ones I labelled failed, basically the lag time fitted ones,  On the Theta table there seems to be v little information presented. Did I run them correctly?  Please see image as an example

 

Attached Files



#7 Simon Davis

Simon Davis

    Advanced Member

  • Administrators
  • 1,328 posts

Posted 16 May 2019 - 09:27 AM

Sometimes i find that the results match exacly my intial estimates i.e. the algorithm didn't start optimising them.  I often find tweaking them very slightly and/or adding lower bounds of zero is enough to get results upon re-execution.  I'm guessing the starting conditions are just a little off for the algorithm to get to work.

Oh and when putting MDV flag back in so time zero conc is unknown, a  lag still seemed to help.


Edited by Simon Davis, 16 May 2019 - 09:36 AM.





0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users