mlr_pipeops_responsecompose: PipeOpResponseCompositor

In mlr-org/mlr3proba: Probabilistic Supervised Learning for 'mlr3'

PipeOpResponseCompositor

Description

Uses a predicted survival distribution (distr) in a PredictionSurv to estimate (or 'compose') an expected survival time (response) prediction. Practically, this PipeOp summarizes an observation's survival curve/distribution to a single number which can be either the restricted mean survival time or the median survival time.

Dictionary

This PipeOp can be instantiated via the dictionary mlr3pipelines::mlr_pipeops or with the associated sugar function mlr3pipelines::po():

PipeOpResponseCompositor$new()
mlr_pipeops$get("responsecompose")
po("responsecompose")

Input and Output Channels

PipeOpResponseCompositor has one input channel named "input", which takes NULL during training and PredictionSurv during prediction.

PipeOpResponseCompositor has one output channel named "output", producing NULL during training and a PredictionSurv during prediction.

The output during prediction is the PredictionSurv from the input but with the response predict type overwritten by the given method.

State

The ⁠$state⁠ is left empty (list()).

Parameters

• method :: character(1)
  Determines what method should be used to produce a survival time (response) from the survival distribution. Available methods are "rmst" and "median", corresponding to the restricted mean survival time and the median survival time respectively.

• tau :: numeric(1)
  Determines the time point up to which we calculate the restricted mean survival time (works only for the "rmst" method). If NULL (default), all the available time points in the predicted survival distribution will be used.

• add_crank :: logical(1)
  If TRUE then crank predict type will be set as -response (as higher survival times correspond to lower risk). Works only if overwrite is TRUE.

• overwrite :: logical(1)
  If FALSE (default) and the prediction already has a response prediction, then the compositor returns the input prediction unchanged. If TRUE, then the response (and the crank, if add_crank is TRUE) will be overwritten.

Internals

The restricted mean survival time is the default/preferred method and is calculated as follows:

T_{i,rmst} \approx \sum_{t_j \in [0,\tau]} (t_j - t_{j-1}) S_i(t_j)

where T is the expected survival time, \tau is the time cutoff/horizon and S_i(t_j) are the predicted survival probabilities of observation i for all the t_j time points.

The T_{i,median} survival time is just the first time point for which the survival probability is less than 0.5. If no such time point exists (e.g. when the survival distribution is not proper due to high censoring) we return the last time point. This is not a good estimate to use in general, only a reasonable substitute in such cases.

Super class

mlr3pipelines::PipeOp -> PipeOpResponseCompositor

Methods

Public methods

• PipeOpResponseCompositor$new()

• PipeOpResponseCompositor$clone()

Inherited methods

• mlr3pipelines::PipeOp$help()
• mlr3pipelines::PipeOp$predict()
• mlr3pipelines::PipeOp$print()
• mlr3pipelines::PipeOp$train()

---

Method new()

Creates a new instance of this R6 class.

Usage

PipeOpResponseCompositor$new(id = "responsecompose", param_vals = list())

Arguments

id

(character(1))
Identifier of the resulting object.

param_vals

(list())
List of hyperparameter settings, overwriting the hyperparameter settings that would otherwise be set during construction.

---

Method clone()

The objects of this class are cloneable with this method.

Usage

PipeOpResponseCompositor$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

References

Zhao, Lihui, Claggett, Brian, Tian, Lu, Uno, Hajime, Pfeffer, A. M, Solomon, D. S, Trippa, Lorenzo, Wei, J. L (2016). “On the restricted mean survival time curve in survival analysis.” Biometrics, 72(1), 215–221. ISSN 1541-0420, \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1111/BIOM.12384")}, https://onlinelibrary.wiley.com/doi/full/10.1111/biom.12384.

See Also

pipeline_responsecompositor

Other survival compositors: mlr_pipeops_compose_breslow_distr, mlr_pipeops_crankcompose, mlr_pipeops_distrcompose

Examples

## Not run: 
  library(mlr3pipelines)
  task = tsk("rats")

  # add survival time prediction type to the predictions of a Cox model
  # Median survival time as response
  pred = lrn("surv.coxph")$train(task)$predict(task)
  por = po("responsecompose", param_vals = list(method = "median", overwrite = TRUE))
  por$predict(list(pred))[[1L]]
  # mostly improper survival distributions, "median" sets the survival time
  # to the last time point

  # RMST (default) as response, while also changing the `crank` to `-response`
  por = po("responsecompose", param_vals = list(overwrite = TRUE, add_crank = TRUE))
  por$predict(list(pred))[[1L]]

## End(Not run)

mlr-org/mlr3proba documentation built on April 12, 2025, 4:38 p.m.