Random Survival Forest for Recurrent Events.

recforest

{recforest} offers a flexible solution for analyzing recurrent events in survival data, outperforming traditional methods like the Cox model, which struggle with repeated events (e.g., hospital readmissions) and terminal events like death. By leveraging machine learning (Random Survival Forests), RecForest models both the timing and frequency of events, even with right-censored data, leading to more accurate predictions and insights, ultimately aiding in better decision-making and patient care.

The methodology is fully described in Murris, J., Bouaziz, O., Jakubczak, M., Katsahian, S., & Lavenu, A. (2024).

Overview

Installation

You can install the development version of {recforest} like so:

TODO : TO BE COMPLETED ONCE THE PACKAGE IS ON GITHUB / CRAN

# Install from CRAN:
install.packages("recforest")

# Or the development version from GitHub:
# install.packages("pak")
pak::pak("XXXX/recforest")

Usage

A example dataset is provided with the package. It is a modified version of the bladder1 dataset, studying bladder cancer recurrences, from the {survival} package, adapted to be usable with {recforest}. Please use ?survival::bladder and ?bladder1_recforest to get more information about the dataset.

library(recforest)
data("bladder1_recforest")
head(bladder1_recforest)
#> # A tibble: 6 × 8
#>      id t.start t.stop treatment number  size death event
#>   <int>   <int>  <int> <fct>      <int> <int> <dbl> <dbl>
#> 1     1       0      0 placebo        1     1     0     0
#> 2     2       0      1 placebo        1     3     0     0
#> 3     3       0      4 placebo        2     1     0     0
#> 4     4       0      7 placebo        1     1     0     0
#> 5     5       0     10 placebo        5     1     0     0
#> 6     6       0      6 placebo        4     1     0     1

Train the model

trained_forest <- train_forest(
  data = bladder1_recforest,
  id_var = "id",
  covariates = c("treatment", "number", "size"),
  time_vars = c("t.start", "t.stop"),
  death_var = "death",
  event = "event",
  n_trees = 5,
  n_bootstrap = 70,
  mtry = 2,
  minsplit = 3,
  nodesize = 15,
  method = "NAa",
  min_score = 5,
  max_nodes = 20,
  seed = 111,
  parallel = FALSE,
  verbose = FALSE
)

A full explanation of the data-related and model-related parameters is provided in the vignette (see Further details).

Using parallel computing

The implementation of parallel computing in this package is based on the {future} and {future.apply} packages. To enable parallel processing, the parallel parameter must be set to TRUE. The number of cores to use can be specified by adjusting the workers parameter in the future::plan() function, which configures the parallelization strategy. Two commonly used strategies are:

• Multicore: Recommended for UNIX systems, this strategy uses multiple distinct processes on the available cores. Example of implementation:

  future::plan(future::multicore, workers = n_cores - 1)
  

• Multisession: Suitable for Windows systems or for more isolated executions, it launches multiple R sessions. Example of implementation:

  future::plan(future::multisession, workers = n_cores - 1)
  

In both cases, the number of cores (n_cores) should be defined by the user based on the available resources.

The following can be run before training the model:

# Define the strategy and number of cores
n_cores <- min(future::availableCores(), n_trees)
future::plan(future::multisession, workers = n_cores - 1)

If you wish to use parallel computing, please refer to the {future}package documentation for more information.

Analyzing results

print(trained_forest)
#>
#> ── Tree 1 ──
#>
#> ℹ Number of nodes : 5
#> ℹ c_index : 0.7490882567469
#> ℹ mse_imse : 316.52338769398
#> ℹ mse_iscore : -23.8459208072833
#>
#> ── Tree 2 ──
#>
#> ℹ Number of nodes : 7
#> ℹ c_index : 0.749320446994866
#> ℹ mse_imse : 317.020277729943
#> ℹ mse_iscore : -24.3379911330769
#>
#> ── Tree 3 ──
#>
#> ℹ Number of nodes : 9
#> ℹ c_index : 0.725611597704621
#> ℹ mse_imse : 552.713717581106
#> ℹ mse_iscore : -260.02620872387
#>
#> ── Tree 4 ──
#>
#> ℹ Number of nodes : 5
#> ℹ c_index : 0.757197981596913
#> ℹ mse_imse : 391.055121586804
#> ℹ mse_iscore : -98.20144371709
#>
#> ── Tree 5 ──
#>
#> ℹ Number of nodes : 7
#> ℹ c_index : 0.75103734439834
#> ℹ mse_imse : 419.451806839106
#> ℹ mse_iscore : -126.772535366445

summary(trained_forest)
#>
#> ── Data summary ────────────────────────────────────────────────────────────────
#> ℹ Number of individuals : 118
#> ℹ Number of predictors : 3
#>
#> ── Model parameters ────────────────────────────────────────────────────────────
#> ℹ mtry : 2
#> ℹ minsplit : 3
#> ℹ nodesize : 15
#> ℹ method : NAa
#> ℹ min_score : 5
#> ℹ max_nodes : 20
#>
#> ── Metrics ─────────────────────────────────────────────────────────────────────
#> ℹ c_index : 0.746451125488328
#> ℹ mse_imse : 399.352862286188
#> ℹ mse_iscore : -106.636819949553
#> ℹ computation time (seconds) : 4.4

Prediction

The model can be used to predict an expected mean cumulative number of recurrent events per individual at the end of follow-up.

predictions <- predict(
  trained_forest,
  newdata = bladder1_recforest,
  id_var = "id",
  covariates = c("treatment", "number", "size"),
  time_vars = c("t.start", "t.stop"),
  death_var = "death"
)

Further details

A deeper explanation of the methodology and the features of the package can be found in the Vignettes.

The Vignettes are structured as follows:

• How I can predict events on a new dataset?
• How can I get more details about the methodology?: A detailed explanation of the methodology used in the package.
• How can I assess the influence of explanatory variables on the event? (To be expected soon).