Machine Learning Feature Selection for High Dimensional Survival Data

A unified, flexible framework for high dimensional feature selection in the presence of a survival outcome. Provides multiple machine learning approaches (Cox elastic net, random survival forest, accelerated oblique random survival forest, gradient-boosted Cox, stability selection, classical univariate Cox screening, pseudo- observation bridging to arbitrary regression learners, and Fine-Gray competing risks selection) under a single interface. Adds causal survival forest estimation of heterogeneous treatment effects on survival (experimental), conformal survival prediction with finite- sample coverage guarantees, and time-dependent 'SHAP' explanations via 'SurvSHAP(t)'. Methodology is based on regularised Cox regression (2011) , random survival forests (2008) , oblique random survival forests (2024) , stability selection (2010) , causal survival forests (2023) , time-dependent survival explanations (2023) , conformal survival prediction (2023) , the Fine-Gray model for competing risks (1999) , and pseudo-observation regression (2010) .

highMLR 1.0.0

Machine learning feature selection for high dimensional survival data.

What's new in 1.0.0

Complete rewrite, superseding the original CRAN release v0.1.1. v0.1.1 offered six hardcoded functions, each fitting a single Cox-based model in a feature-by-feature loop. This version replaces that with a single, flexible entry point that dispatches to modern ML backends:

Plus companions: highmlr_compare(), highmlr_stability(), highmlr_explain() (time-dependent SHAP via survex), highmlr_screen(), highmlr_report(), and the additional tools highmlr_causal() (causal survival forest via grf, experimental) and highmlr_conformal() (conformal survival prediction intervals).

Install (development)

# install.packages("remotes")
remotes::install_local("path/to/highMLR")

Quick start

library(highMLR)
data(hnscc)

# One-liner: Cox elastic net with 5-fold CV
fit <- highmlr(hnscc, time = "OS", status = "Death", method = "coxnet")
print(fit)
plot(fit)

# Compare three methods
cmp <- highmlr_compare(hnscc, "OS", "Death",
                       methods = c("coxnet", "rsf", "xgboost"))
cmp$summary

Breaking changes from 0.1.1

All v0.1.1 functions (mlhighCox, mlhighKap, mlhighFrail, mlhighHet, mlclassCox, mlclassKap) are removed. Use highmlr() with the appropriate method argument. The univariate method reproduces the v0.1.x screening behaviour.