topepo / survnip

Parsnip Wrappers for Survival Models

survnip

survnip (survival + parsnip) is a “parsnip-adjacent” packages with model definitions for survival analysis packages.

Installation

This package is still in early development. You need to install the developmental branch of parsnip as well.

# install.packages("devtools")
remotes::install_github("tidymodels/parsnip#396")
devtools::install_github("EmilHvitfeldt/survnip")

Prediction Types

The addition of censored regression comes with changes. One of these changes is the quantities we would like to predict from the model. The 3 quantities we will consider are: "time", "survival", and "linear_pred".

To showcase these the differences, here is a simple Cox regression model fitted on the lung data set.

library(survnip)
#> Loading required package: parsnip
library(survival)

cox_mod <-
  cox_reg() %>%
  set_engine("survival") %>%
  fit(Surv(time, status) ~ age + ph.ecog, data = lung)

cox_mod
#> parsnip model object
#> 
#> Fit time:  18ms 
#> Call:
#> survival::coxph(formula = Surv(time, status) ~ age + ph.ecog, 
#>     data = data, x = TRUE)
#> 
#>             coef exp(coef) se(coef)     z        p
#> age     0.011281  1.011345 0.009319 1.211 0.226082
#> ph.ecog 0.443485  1.558128 0.115831 3.829 0.000129
#> 
#> Likelihood ratio test=19.06  on 2 df, p=7.279e-05
#> n= 227, number of events= 164 
#>    (1 observation deleted due to missingness)

time

when we specify type = "time" then we get back the predicted survival time of an observation based on its predictors. The survival time is the time it takes for the observation to observe an event.

predict(cox_mod, type = "time", new_data = lung)
#> # A tibble: 228 x 1
#>    .pred_time
#>         <dbl>
#>  1       342.
#>  2       474.
#>  3       511.
#>  4       389.
#>  5       498.
#>  6       342.
#>  7       263.
#>  8       256.
#>  9       401.
#> 10       278.
#> # … with 218 more rows

Here we see that the first patient is predicted to have 342.355592 days left.

survival

when we specify type = "survival" then we are trying to get the properbility of survival (not observing an event) at a given time .time.

pred_vals_survival <- predict(cox_mod, 
                              type = "survival", 
                              new_data = lung, 
                              .time = c(100, 200))

pred_vals_survival
#> # A tibble: 228 x 1
#>        .pred_survival
#>    <list<tbl_df[,2]>>
#>  1            [2 × 2]
#>  2            [2 × 2]
#>  3            [2 × 2]
#>  4            [2 × 2]
#>  5            [2 × 2]
#>  6            [2 × 2]
#>  7            [2 × 2]
#>  8            [2 × 2]
#>  9            [2 × 2]
#> 10            [2 × 2]
#> # … with 218 more rows

pred_vals_survival$.pred_survival[[1]]
#> # A tibble: 2 x 2
#>   .time .pred_survival
#>   <chr>          <dbl>
#> 1 100            0.855
#> 2 200            0.649

here we see that the first patient has a 85.5% probability of survival after 100 days and 64.9% probability of survival after 200 days.

linear_pred

when we specify type = "linear_pred" then we get back the linear predictor for the observation according to the model.

predict(cox_mod, type = "linear_pred", new_data = lung)
#> # A tibble: 228 x 1
#>    .pred_linear_pred
#>                <dbl>
#>  1            0.152 
#>  2           -0.359 
#>  3           -0.495 
#>  4           -0.0401
#>  5           -0.450 
#>  6            0.152 
#>  7            0.527 
#>  8            0.561 
#>  9           -0.0852
#> 10            0.449 
#> # … with 218 more rows

here we see that the linear predictor of the first observation is 0.1517.

Prediction type table

alias	engine	survival	linear_pred	time
boost_tree	mboost	TRUE	TRUE	FALSE
decision_tree	rpart	TRUE	FALSE	TRUE
decision_tree	party	TRUE	FALSE	TRUE
rand_forest	party	TRUE	FALSE	TRUE
bag_tree	ipred	TRUE	FALSE	TRUE
cox_reg	survival	TRUE	TRUE	TRUE
cox_reg	glmnet	TRUE	TRUE	FALSE