Support for time-dependent variables and [start, stop) format

[zhiweill]

In survival package, the most common way to encode time-dependent covariates is to use the (start, stop] form
of the model: Surv(time1, time2, status). Below is what the timedep vignettes of survival package said:

In mlr3, as_task_surv() do have parameters time2, for example, as_task_surv(mydata, time = "time1",time2 = "time2", event = "status"). However, when I run the codes below, the selected feature included the time2.

lrn = lrn("surv.cv_glmnet")
lrn$train(mytask, split$train)
lrn$selected_features()

Does it mean the surv learners like glmnet regard the time2 as the predictor instead of the response variable Surv(...)? What survival models in mlr3 support Surv(time1, time2, status) encoding?

Thanks!!!

[bblodfon]

Hi! We do not support anything other than right-censored survival tasks atm. So yes, I think right now time2 is used as a variable in surv.cv_glmnet via mlr3. Changes are needed in compatible learners to understand this time-dependent format even though the formulation of the SurvTask in mlr3proba works atm.

To implement this I need either toooooooo much free time :) or a research student to help! In this issue I have more details about how this should look like (different types of survival tasks, etc.)

[zhiweill]

Thanks for your reply :)

[agalecki]

Hi, It appears it works now. Just add type = "counting" argument in your task definition

library(mlr3)
library(mlr3proba)
mytask = as_task_surv(survival::heart, time = "start", time2 = "stop", event = "event", type="counting")
mytask

<TaskSurv:survival::heart> (172 x 8)
* Target: start, stop, event
* Properties: -
* Features (5):
  - dbl (4): age, id, surgery, year
  - fct (1): transplant
> 

[zhiweill]

Thank you for the note :)