Package 'survnet'

Create binary response matrix for survival data

Description

Creates matrix with at-risk and event information. Format: (S_1, ..., S_K, E_1, ..., E_K). Dimensions: obs X 2*causes*time.

Usage

convert_surv_cens(time, status, breaks, num_causes)

Arguments

time	Survival time
status	Censoring indicator: 0 for censored observations, positive values for events.
breaks	Right interval limits for discrete survival time.
num_causes	Number of competing risks.

Value

Binary response matrix.

---

Cumulative incidence log-likelihood

Description

Likelihood of parametric inference for the cumulative incidence functions as defined by Jeong & Fine 2006. Also used by Lee et al. 2018.

Usage

loss_cif_loglik(num_intervals, num_causes = 1)

Arguments

num_intervals	Number of time intervals
num_causes	Number of causes for competing risks

Details

Data structure:

y_true True survival: Matrix with at-risk and event information. Format: (S_1, ..., S_K, E_1, ..., E_K). Dimensions: obs X 2*causes*time.

y_pred Network output: One probability for each time and cause. Format: (y_11, ..., y_T1, ..., y_TK). Dimensions: obs X causes*time.

Value

Negative log-likelihood

References

• Jeong, J. & Fine, J. (2006). Direct parametric inference for the cumulative incidence function. J R Stat Soc Ser C Appl Stat 55:187-200. https://doi.org/10.1111/j.1467-9876.2006.00532.x.

• Lee, C., Zame, W.R., Yoon, J. & van der Shaar, M. (2018). DeepHit: A deep learning approach to survival analysis with competing risks. AAAI 2018. http://medianetlab.ee.ucla.edu/papers/AAAI_2018_DeepHit.

---

survnet prediction

Description

survnet prediction

Usage

## S3 method for class 'survnet'
predict(object, newdata, cause = NULL, ...)

Arguments

object	survnet object
newdata	New data predictors: matrix, array or data.frame.
cause	Select cause for competing risks, NULL returns list of all causes.
...	Further arguments passed to or from other methods.

Value

Cumulative incidence function of selected or all causes.

---

Artificial neural networks for survival analysis

Description

Artificial neural networks for survival analysis

Usage

survnet(y, x, breaks, units = c(3, 5), units_rnn = c(4, 6),
  units_causes = c(3, 2), epochs = 100, batch_size = 16,
  validation_split = 0.2, loss = loss_cif_loglik,
  activation = "tanh", rnn_type = "LSTM", skip = TRUE,
  dropout = rep(0, length(units)), dropout_rnn = rep(0,
  length(units_rnn)), dropout_causes = rep(0, length(units_causes)),
  l2 = rep(0, length(units)), l2_rnn = rep(0, length(units_rnn)),
  l2_causes = rep(0, length(units_causes)),
  optimizer = optimizer_rmsprop(lr = 0.001), verbose = 2)

Arguments

y	Survival outcome: matrix, data.frame or Surv() object.
x	Predictors: matrix, data.frame or array (time-series). Also accepts a list of matrix/data.frame and array for both time-constant and time-series predictors.
breaks	Right interval limits for discrete survival time.
units	Vector of units, each specifying the number of units in one hidden layer.
units_rnn	Vector of units for recurrent layers.
units_causes	Vector of units for cause-specific layers (competing risks only). Either a vector (will be repeated for each cause) or a list of vectors with layers for each cause.
epochs	Number of epochs to train the model.
batch_size	Number of samples per gradient update.
validation_split	Fraction in [0,1] of the training data to be used as validation data.
loss	Loss function.
activation	Activation function.
rnn_type	Type of RNN layers. Either "LSTM" (default), "GRU", "CUDNN_LSTM" or "CUDNN_GRU".
skip	Add skip connection from input and RNN layers to cause-specific layers.
dropout	Vector of dropout rates after each hidden layer. Use 0 for no dropout (default).
dropout_rnn	Vector of dropout rates after each recurrent layer. Use 0 for no dropout (default).
dropout_causes	Vector of dropout rates after each cause-specific layer. Use 0 for no dropout (default).
l2	Vector of L2 regularization factors for each hidden layer. Use 0 for no regularization (default).
l2_rnn	Vector of L2 regularization factors for each recurrent layer. Use 0 for no regularization (default).
l2_causes	Vector of L2 regularization factors for each cause-specific layer. Use 0 for no regularization (default).
optimizer	Name of optimizer or optimizer instance.
verbose	Verbosity mode (0 = silent, 1 = progress bar, 2 = one line per epoch).

Value

Fitted model.

Examples

library(survival)
library(survnet)

# Survival data
y <- veteran[, c(3, 4)]
x <- veteran[, c(-2, -3, -4)]
x <- data.frame(lapply(x, scale))
breaks <- c(1, 50, 100, 200, 500, 1000)

# Fit simple model
fit <- survnet(y = y, x = x, breaks = breaks)
plot(fit$history)

---