Changes in version 1.0.1 (2026-05-23)                  

  - Default feature set in highmlr() now excludes non-numeric columns
    (e.g. identifiers) automatically, preventing an error when features
    is left unspecified.
  - The coxnet backend now validates that supplied features are numeric
    and reports a clear error otherwise.
  - Learner backends (rsf, aorsf, xgboost) now default to at most two
    threads, configurable via options(), to respect multi-core policies
    in batch-checking environments.

                        Changes in version 1.0.0                        

Major rewrite. This release supersedes the original CRAN version 0.1.1
and is not backward-compatible with it. The previous six functions
(mlhighCox, mlhighKap, mlhighFrail, mlhighHet, mlclassCox, mlclassKap)
and their mlr3/coxme/missForest backend have been replaced by a single,
unified interface. Code written for 0.1.1 will not run unchanged.

New unified interface

  - Single highmlr() entry point dispatching to eight feature-selection
    methods, all returning a common highmlr_fit S3 object: Cox elastic
    net (coxnet), random survival forest (rsf), accelerated oblique RSF
    (aorsf), gradient-boosted Cox (xgboost), stability selection
    (stability), univariate Cox screening (univariate),
    pseudo-observation bridging (pseudo), and Fine-Gray competing-risks
    selection (finegray).
  - Common verbs for the fit object: print(), summary(), plot(), coef(),
    predict().
  - Companion functions: highmlr_compare(), highmlr_stability(),
    highmlr_explain(), highmlr_screen(), highmlr_report().

Feature-selection methods

  - method = "pseudo" -- Andersen & Perme (2010) pseudo-observation
    bridge. Computes jackknife pseudo-values for S(t) at chosen
    evaluation times, then fits any regression learner (ranger, xgboost,
    glmnet) on the transformed outcome, allowing arbitrary regression ML
    on right-censored data without the proportional hazards assumption.
  - method = "finegray" -- Fine-Gray subdistribution hazard feature
    selection for competing-risks data. Status coded 0/1/2+ (cause of
    interest configurable). Returns subdistribution hazard ratios with
    FDR-adjusted p-values.

Causal, conformal, and explanation tools

  - highmlr_causal() -- EXPERIMENTAL. Causal survival forest via grf for
    heterogeneous treatment effects on survival. Returns per-patient
    CATE estimates (RMST or survival-probability difference) with
    standard errors, the average treatment effect, and covariate
    importance.
  - highmlr_conformal() -- conformal prediction intervals for survival
    time with finite-sample marginal coverage (Candes, Lei & Ren, 2023),
    using inverse-probability-of-censoring weighting on the calibration
    set.
  - highmlr_explain() -- time-dependent SHAP via SurvSHAP(t) (Krzyzinski
    et al., 2023) by default, returning SHAP values per feature, per
    patient, across time, with aggregated mean |SHAP| importance.
    Permutation and break-down methods are also available.

Dependencies

  - Imports: survival, glmnet, ranger, aorsf, xgboost, stabs, survex,
    grf, prodlim, cmprsk, future, future.apply, tibble, ggplot2, rlang,
    stats, utils.
  - Suggests: knitr, rmarkdown, testthat, mice, riskRegression.
  - The mlr3, mlr3learners, gtools, dplyr, coxme, missForest, and R6
    dependencies from 0.1.1 are no longer used.

Notes

  - Neural network methods (DeepSurv, DeepHit) are deferred to a future
    release to keep this version free of torch/Python dependencies.
  - Datasets hnscc and srdata are retained from 0.1.1, now with expanded
    and corrected documentation.