[LRNRQ]: Add cmprsk.crr learner from package cmprsk

[agalecki]

Algorithm

Fine-Gray model for subdistribution functions in competing risks

Package

cmprsk

Supported types

• classif
• clust
• dens
• regr
• cmprsk

I have checked that this is not already implemented in

• mlr3
• mlr3learners
• mlr3extralearners
• Other core packages (e.g. mlr3proba, mlr3keras)

Why do I think this is a useful learner?

cmprsk::crr is considered to be a primary choice for fitting the F-G model

Further Optional Comments

A working version of the learner is available
here. I appreciate your consideration.

[agalecki]

Hello,

I’ve noticed that the .predict() method in LearnerCompRisksAalenJohansen (from R/LearnerCompRisksAalenJohansen.R in the competing_risk branch, PR #433) returns a list of CIFs, one per competing event. Is this (conceptually sound) behavior a fixed requirement of the learner or the mlr3proba competing risks framework? I’m asking because, in many use cases, it’s common to designate one ‘main event’ and compute a single CIF for that event, adjusted for all remaining competing risks (e.g., return a list with one component). Could the learner be adapted to support this simplified output instead of returning separate CIFs for each competing event? Many thanks for considering my request.

[bblodfon]

Hi @agalecki,

Yes, this is more a requirement from the framework, the code checks for this (see here and here). The idea is that if a model can return a CIF for a specific competing event, then it can for all other causes as well (at least I haven't find a case where this is impossible!).

Then, instead of making the learner cause-specific as you suggest, I think it's better to apply cause-specific evaluation. In the competing risks PR I have implemented the AUC(t) from the riskRegression package (see here) which can be used to evaluate discrimination performance at a specific time point and either for a specific cause or average across all causes.

[bblodfon]

Dear Andrzej (@agalecki),

Just a quick note regarding your PR — the latest mlr3proba now supports right-censored competing risks tasks. It would be great to revisit the Fine-Gray model integration, even with the caveats we discussed.

Do you think you’ll have time to update the code to work with the new version? It should mostly involve simplifying for right-censoring and ensuring all CIFs are returned I think?

BR, John

[agalecki]

I've revised the cmprsk.crr learner (version 0.3.0) with significant updates. I'll open a new issue to reflect these changes and provide a fresh submission. Please let me know if I should continue here instead. Thank you!

[bblodfon]

Whatever suits you - opening a new PR and removing the old one is super fine as well!

[agalecki]

I made some progress, but I have to admit it is more difficult than I thought. When I run an example (see below), I am getting an error message related to mirai encapsulation. Please advise how to address this issue.

> options(mlr3crr.debug = FALSE)
> 
> # Run from a folder containing source of mlr3extralearners 
> library(mlr3)
> rm(list =ls())
> 
> devtools::load_all()
ℹ Loading mlr3extralearners
> 
> 
> task = tsk("pbc")
> sel_feat =c("age", "bili","sex")
> 
> task$select(sel_feat)
> task

── <TaskCompRisks> (276x5): Primary Biliary Cholangitis ──────────────────────────────────────────────────────────────────────────
• Target: time and status
• Properties: -
• Features (3):
  • int (1): age
  • dbl (1): bili
  • fct (1): sex
> 
> ids = partition(task)
> learner = lrn("cmprsk.crr")
> learner$train(task, ids$train)
> print(learner$model)
$`1`
convergence:  TRUE 
coefficients:
     age     bili      sex 
-0.11850  0.03734 -1.61200 
standard errors:
[1] 0.02510 0.02851 0.52790
two-sided p-values:
    age    bili     sex 
2.3e-06 1.9e-01 2.3e-03 

$`2`
convergence:  TRUE 
coefficients:
    age    bili     sex 
0.04713 0.15140 0.10770 
standard errors:
[1] 0.01232 0.01865 0.39040
two-sided p-values:
    age    bili     sex 
1.3e-04 4.4e-16 7.8e-01 

> preds = learner$predict(task, ids$test)
> preds$score()
cmprsk.auc 
 0.8331487 
> 
> 
> # From run_experiment()
> learner_encapsulated = learner$clone(deep=TRUE)
> 
> # resample works with mirai encapsulation skipped
> rr0 =  resample(task, learner_encapsulated, rsmp("holdout"), store_models = TRUE) 
INFO  [13:36:56.440] [mlr3] Resampling 'holdout' is being instantiated on task 'pbc'
INFO  [13:36:56.469] [mlr3] Applying learner 'cmprsk.crr' on task 'pbc' (iter 1/1)
> log0 = rr0$learners[[1]]$state$log
> log0
Empty data.table (0 rows and 3 cols): stage,class,msg
> 
> # resample does not work with mirai encapsulation 
> learner_encapsulated$encapsulate("mirai", default_fallback(learner_encapsulated))
> rr =  resample(task, learner_encapsulated, rsmp("holdout"), store_models = TRUE)
INFO  [13:36:56.928] [mlr3] Resampling 'holdout' is being instantiated on task 'pbc'
INFO  [13:36:56.949] [mlr3] Applying learner 'cmprsk.crr' on task 'pbc' (iter 1/1)
INFO  [13:37:00.336] [mlr3] Calling train method of fallback 'cmprsk.aalen' on task 'pbc' with 184 observations {learner: <LearnerCompRisksAalenJohansen/LearnerCompRisks/Learner/R6>}
ERROR [13:37:03.739] [mlr3] predict: no terms component nor attribute
>  
> log = rr$learners[[1]]$state$log
> log
     stage  class                                    msg
    <fctr>  <ord>                                 <char>
1: predict  error       no terms component nor attribute
2: predict output Using fallback learner for predictions
> 

[bblodfon]

Thanks for the effort! I would try to open a new PR with the learner code + including the automated tests so that we can check this more thoroughly. I can't guess what is wrong from just seeing the output you sent me!

Note there were some things related to encapsulation and default fallback learners that are very technical and mostly were solved in the newest mlr3 and mlr3proba versions. Also, I am porting everything related to competing risks to a new package called mlr3cmprsk (see PR), which I will make public soon (and at some point will be on CRAN).