Model-Agnostic Survival Analysis in R
survivalisttoo provides a flexible framework for performing survival analysis using any machine learning algorithm.
✨ Features
- Model-agnostic survival modeling
- Plug-and-play with arbitrary regression/ML learners
- Simple
predict()interface
🚀 Quick Example
library(survivalisttoo)
library(glmnet)
library(survival)
data(pbc)
pbc2 <- pbc[!is.na(pbc$trt), ]
pbc2$event <- as.integer(pbc$status[!is.na(pbc$trt)] == 2)
pbc2$sex_n <- as.integer(pbc2$sex == "f")
feat_cols <- c("trt","age","sex_n","ascites","hepato","spiders","edema",
"bili","chol","albumin","copper","alk.phos","ast",
"trig","platelet","protime","stage")
df <- pbc2[, c("time", "event", feat_cols)]
# Simple imputation
for (col in feat_cols)
if (any(is.na(df[[col]])))
df[[col]][is.na(df[[col]])] <- median(df[[col]], na.rm = TRUE)
set.seed(42)
idx_train <- sample(nrow(df), floor(0.75 * nrow(df)))
train <- df[idx_train, ]
test <- df[-idx_train, ]
X_tr <- as.matrix(train[, feat_cols])
X_te <- as.matrix(test[, feat_cols])
# Custom regression model
regr_lm <- function(X, y, ...) {
lm(y ~ ., data = data.frame(X, y = y))
}
# Fit model-agnostic survival boosting
fit_boost_lm <- cox_gradient_boost(
X=X_tr, time=train$time, event=train$event, regr_fun=regr_lm
)
# Classical Cox model (baseline)
fit_cox <- coxph(
Surv(time, event) ~ .,
data = train[, c("time","event",feat_cols)],
x = TRUE
)
# Evaluate
y_te <- Surv(test$time, test$event)
ci_blm <- glmnet::Cindex(predict(fit_boost_lm, X_te), y_te)
ci_cox <- glmnet::Cindex(predict(fit_cox, newdata = test), y_te)
cat("\n=== Test-set C-index ===\n")
cat(sprintf(" CoxBoost (LM): %.4f\n", ci_blm))
cat(sprintf(" Classical Cox : %.4f\n", ci_cox))🔌 Using Different ML Models
You can plug in different learners via wrappers (e.g. from mlS3):
library(mlS3)
require(glmnet)
require(survival)
X_tr <- train[, feat_cols]
X_te <- test[, feat_cols]
y_te <- Surv(test$time, test$event)
# Elastic net (glmnet)
fit_glmnet <- cox_gradient_boost(
X_tr, train$time, train$event,
mlS3::wrap_glmnet,
alpha = 1,
show_progress = FALSE
)
glmnet::Cindex(predict(fit_glmnet, X_te), y_te)
# SVM (radial)
fit_svm <- cox_gradient_boost(
X_tr, train$time, train$event,
mlS3::wrap_svm,
kernel = "radial",
show_progress = FALSE
)
glmnet::Cindex(predict(fit_svm, X_te), y_te)
# caret model
fit_caret <- cox_gradient_boost(
X_tr, train$time, train$event,
mlS3::wrap_caret,
method = "enet",
show_progress = FALSE
)
glmnet::Cindex(predict(fit_caret, X_te), y_te)📚 How it works
For now, the package implements a gradient boosting strategy for survival analysis in cox_gradient_boost:
- Iteratively fits a base learner on pseudo-residuals
- Uses any regression/ML model as the base learner
- Produces risk scores compatible with survival metrics (e.g. C-index)
This enables combining:
- statistical survival modeling
- modern machine learning flexibility