What’s new in 0.21#

scikit-survival 0.21.0 (2023-06-11)#

This is a major release bringing new features and performance improvements.

sksurv.nonparametric.kaplan_meier_estimator() can estimate pointwise confidence intervals by specifying the conf_type parameter.
sksurv.ensemble.GradientBoostingSurvivalAnalysis supports early-stopping via the monitor parameter of sksurv.ensemble.GradientBoostingSurvivalAnalysis.fit().
sksurv.metrics.concordance_index_censored() has a significantly reduced memory footprint. Memory usage now scales linear, instead of quadratic, in the number of samples.
Fitting of sksurv.tree.SurvivalTree, sksurv.ensemble.RandomSurvivalForest, or sksurv.ensemble.ExtraSurvivalTrees is about 3x faster.
Finally, the release adds support for Python 3.11 and pandas 2.0.

Fix bug where times passed to sksurv.metrics.brier_score() was downcast, resulting in a loss of precision that may lead to duplicate time points (#349).
Fix inconsistent behavior of evaluating functions returned by predict_cumulative_hazard_function or predict_survival_function (#375).

sksurv.nonparametric.kaplan_meier_estimator() and sksurv.nonparametric.CensoringDistributionEstimator support returning confidence intervals by specifying the conf_type parameter (#348).
Configure package via pyproject.toml (#347).
Add support for Python 3.11 (#350).
Add support for early-stopping to sksurv.ensemble.GradientBoostingSurvivalAnalysis (#354).
Do not use deprecated pkg_resources API (#353).
Significantly reduce memory usage of sksurv.metrics.concordance_index_censored() (#362).
Set criterion attribute in sksurv.tree.SurvivalTree such that sklearn.tree.plot_tree() can be used (#366).
Significantly improve speed to fit a sksurv.tree.SurvivalTree, sksurv.ensemble.RandomSurvivalForest, or sksurv.ensemble.ExtraSurvivalTrees (#371).
Expose _predict_risk_score attribute in sklearn.pipeline.Pipeline if the final estimator of the pipeline has such property (#374).
Add support for pandas 2.0 (#373).

Fix wrong number of selected features in the guide Introduction to Survival Analysis (#345).
Fix broken links with nbsphinx 0.9.2 (#367).

The attribute event_times_ of estimators has been replaced by unique_times_ to clarify that these are all the unique times points, not just the once where an event occurred (#371).
Functions returned by predict_cumulative_hazard_function and predict_survival_function of sksurv.tree.SurvivalTree, sksurv.ensemble.RandomSurvivalForest, and sksurv.ensemble.ExtraSurvivalTrees are over all unique time points passed as training data, instead of all unique time points where events occurred (#371).
Evaluating a function returned by predict_cumulative_hazard_function or predict_survival_function will no longer raise an exception if the specified time point is smaller than the smallest time point observed during training. Instead, the value at StepFunction.x[0] will be returned (#375).