BestDose requests

[mhovd]

@R-garreau asked for the following:

• Separate functions from Pmetrics, so that Pmetrics doesn't have to be loaded in full
• Access to posterior parameter values and probabilities for past data
• Option to disable logs for BestDose
• Access to "raw" predictions and their probabilities (both posterior and population?)

Additional Request for Release (2026-02-19)

• would it be possible to define the dose to optimize by -1 ? I use a dose of 0 to add new covariates when its not at the time of an observation or dose.
• Is the interpolation system working ? If not is it possible to have it ?
• Can we remove the uniform prior search ?
  • This is what’s take the longest to run. In the example from where the picture below is the full run was 1.08 min.
  • Breaking it down by section, it was : 5s to recalculate posterior weight and value. 22s to calculate new dose based on posterior weights and 37 second to calculate new dose based the uniform weight.
  • Uniform calculation may be accessible with an option (TRUE or FALSE). But I don't find it useful in clinical practice and run time seem to be twice as long (at least)
• Inclue n_cycle = 0 by default and me be accesible to have the hybrid.

@R-garreau, if you can provide a "wishlist" for the API and how you prefer to work with data and results then we can do larger iterations.

• The ecosystem to read and parse file to to run BestDose is available here LAPKB/BestDoseApp/issues/15
• The way data is handle by the software is available 🔜

[R-garreau]

@mhovd @Siel
Here is the documentation for the hybrid of the legacy BestDose : HYBRID_11-9-10.pdf

[R-garreau]

First real patient (with covariates) Tested with BDcore ! And we have more precise results thant with old BestDose

@mhovd @Siel I've updated the request

[mnneely]

Outstanding!

[mhovd]

Regarding this one

would it be possible to define the dose to optimize by -1 ? I use a dose of 0 to add new covariates when its not at the time of an observation or dose.

You don't need to give a dose of zero to update the covariates. The recommended way is to use a missing observation, or -99 for the observed value.

[mhovd]

Can you clarify this one?

Inclue n_cycle = 0 by default and me be accesible to have the hybrid.

If I understand, you want to run a zero-cycle NPAG run, i.e. only calculate the posterior probabilities? We already have this functionality, so we can export that relatively easily.

[R-garreau]

1- Regarding this one

would it be possible to define the dose to optimize by -1 ? I use a dose of 0 to add new covariates when its not at the time of an observation or dose.

You don't need to give a dose of zero to update the covariates. The recommended way is to use a missing observation, or -99 for the observed value.

So you mean that we can just have a line with evid = 0 and NA in out, and this will suffise to add new covariate value for the patient ?

2 - Can you clarify this one?

Inclue n_cycle = 0 by default and me be accesible to have the hybrid.

If I understand, you want to run a zero-cycle NPAG run, i.e. only calculate the posterior probabilities? We already have this functionality, so we can export that relatively easily.

Yes, it's exaclty that. It seems that some options are not accessible in R (and the cycle number a amongst the non accesible BestDose Options)
It would just be making it accessible when calling the function from R

[mhovd]

So you mean that we can just have a line with evid = 0 and NA in out, and this will suffise to add new covariate value for the patient ?

Yes, exactly!

If I understand, you want to run a zero-cycle NPAG run, i.e. only calculate the posterior probabilities? We already have this functionality, so we can export that relatively easily.

Yes, it's exaclty that. It seems that some options are not accessible in R (and the cycle number a amongst the non accesible BestDose Options) It would just be making it accessible when calling the function from R

This should be available in R, using a regular PM_model$fit(), with algorithm = "MAP" (I think, either that or "POSTPROB". The current framework of BestDose I think can not use it directly, we have to think about how to wrap it nicely.