Experimental DeepHit Learner — surv.deephit • SuperSurv

Fits a single-event discrete-time DeepHit neural network through survivalmodels::deephit(). This optional adapter requires a Python environment containing torch, torchtuples, and pycox and currently supports only uniform observation weights.

Usage

surv.deephit(
  time,
  event,
  X,
  newdata = NULL,
  new.times,
  obsWeights = NULL,
  id = NULL,
  num_nodes = c(32L, 32L),
  activation = "relu",
  batch_norm = TRUE,
  dropout = NULL,
  epochs = 100L,
  batch_size = 128L,
  device = NULL,
  verbose = FALSE,
  seed = 1L,
  cuts = 20L,
  cutpoints = NULL,
  scheme = c("equidistant", "quantiles"),
  mod_alpha = 0.2,
  sigma = 0.1,
  ...
)

Arguments

time: Observed follow-up time.
event: Observed event indicator.
X: Training covariate data frame.
newdata: Covariate data frame used for prediction.
new.times: Times at which survival probabilities are requested.
obsWeights: Optional non-negative observation weights.
id: Currently ignored.
num_nodes: Positive integers giving hidden-layer sizes.
activation: Neural-network activation name.
batch_norm: Whether to use batch normalization.
dropout: Optional dropout probability in [0, 1).
epochs: Number of training epochs.
batch_size: Training and prediction batch size.
device: Optional device passed to survivalmodels.
verbose: Whether the Python backend should print training progress.
seed: Positive integer used for Python, NumPy, and torch random-number generators.
cuts: Number of discrete time intervals, or a vector of cut points supplied through cutpoints.
cutpoints: Optional numeric vector of cut points.
scheme: Cut-point scheme used when cutpoints is NULL.
mod_alpha: Weight placed on the likelihood component of the DeepHit objective.
sigma: Ranking-loss bandwidth.
...: Additional arguments passed to survivalmodels::deephit().

Value

A list with numeric matrix pred and fitted object fit.

Details

Fitted Python objects can be reused in the active R/Python session. Plain saveRDS() is not a portable persistence format for these objects. Native survival curves are mapped to requested times as right-continuous steps; see surv.coxtime() for the boundary convention.

Examples

if (interactive() && requireNamespace("survivalmodels", quietly = TRUE) &&
    requireNamespace("reticulate", quietly = TRUE) &&
    reticulate::py_module_available("pycox")) {
  data("metabric", package = "SuperSurv")
  dat <- metabric[1:60, ]
  X <- dat[, grep("^x", names(dat))[1:3], drop = FALSE]
  fit <- surv.deephit(
    dat$duration, dat$event, X, X[1:4, , drop = FALSE],
    c(50, 100), cuts = 10, epochs = 2, seed = 1
  )
  dim(fit$pred)
}