临床研究生存分析

references/ensemble-models.md

Verified for scikit-survival 0.28.0 on 2026-10-01.

Model families

• SurvivalTree: one log-rank survival tree.
• RandomSurvivalForest: bootstrap-aggregated survival trees with random feature subsets.
• ExtraSurvivalTrees: additional randomization of candidate split thresholds.
• GradientBoostingSurvivalAnalysis: regression-tree gradient boosting.
• ComponentwiseGradientBoostingSurvivalAnalysis: linear componentwise base learners and implicit sparse selection.

Model family does not determine quality in advance. Compare prespecified candidates with identical outer resamples, censoring assumptions, time grids, and preprocessing.

Random survival forest

from sksurv.ensemble import RandomSurvivalForest

model = RandomSurvivalForest(
    n_estimators=500,
    min_samples_split=10,
    min_samples_leaf=8,
    max_features="sqrt",
    n_jobs=1,
    random_state=20260723,
)
model.fit(X_train, y_train)

Current defaults include n_estimators=100, min_samples_split=6, min_samples_leaf=3, max_features="sqrt", bootstrap=True, and low_memory=False.

Each terminal node estimates:

• a survival function using Kaplan-Meier;
• a cumulative hazard function using Nelson-Aalen;
• a risk score summing the estimated cumulative hazard over the training event times; its magnitude is neither a probability nor an expected count at a named horizon.

Forest predictions average tree predictions:

risk = model.predict(X_test)  # (n_test,), higher is riskier
survival = model.predict_survival_function(X_test, return_array=True)
cumulative_hazard = model.predict_cumulative_hazard_function(
    X_test, return_array=True
)
times = model.unique_times_

Array predictions use the model's unique_times_. For an evaluation grid, returned step functions are often more convenient:

import numpy as np

functions = model.predict_survival_function(X_test, return_array=False)
survival_on_grid = np.vstack([fn(evaluation_times) for fn in functions])

Do not treat the numeric magnitude of predict() as an event probability.

Missing values and memory

Current survival trees and forest split logic supports missing values, with fixes aligned to scikit-learn 1.8 in scikit-survival 0.27. This does not remove the need to:

• verify which columns and missingness patterns are supported;
• encode categorical columns consistently;
• keep all learned preprocessing within training folds;
• assess whether missingness itself changes across deployment settings.

low_memory=True reduces stored prediction state but disables survival-function and cumulative-hazard prediction. It is incompatible with Brier-score workflows that require survival probabilities.

OOB estimates

With oob_score=True and bootstrap sampling, oob_score_ provides an internal out-of-bag concordance estimate. It is not a substitute for:

• nested tuning when parameters were selected using OOB results;
• an independent test set;
• censoring-aware probability metrics;
• external validation.

Extra survival trees

from sksurv.ensemble import ExtraSurvivalTrees

model = ExtraSurvivalTrees(
    n_estimators=500,
    min_samples_leaf=8,
    max_features="sqrt",
    n_jobs=1,
    random_state=20260723,
)
model.fit(X_train, y_train)

Extra trees randomize split thresholds in addition to feature selection. They are not guaranteed to be faster, better regularized, or better calibrated for a given dataset. Tune and evaluate them as a distinct candidate under the same protocol.

Permutation importance

Survival forest impurity importance is not implemented as a valid feature_importances_ measure. Use held-out permutation importance with an explicit score:

from sklearn.inspection import permutation_importance

result = permutation_importance(
    fitted_pipeline,
    X_test,
    y_test,
    scoring=None,  # estimator.score: Harrell concordance
    n_repeats=20,
    random_state=20260723,
    n_jobs=1,
)

If Harrell C is not the target, wrap the estimator with the appropriate scikit-survival scorer class before permutation or write a scorer that fits no state on the test set. Importance depends on:

• the metric and horizon;
• correlated features;
• the held-out population;
• preprocessing and random seed.

It is predictive sensitivity, not causal importance.

Tree gradient boosting

from sksurv.ensemble import GradientBoostingSurvivalAnalysis

model = GradientBoostingSurvivalAnalysis(
    loss="coxph",
    learning_rate=0.05,
    n_estimators=300,
    max_depth=2,
    subsample=0.8,
    random_state=20260723,
)
model.fit(X_train, y_train)

Current losses:

• "coxph": Cox partial-likelihood objective; predict() is a higher-is-riskier score, and baseline-based survival/cumulative-hazard methods are available.
• "ipcwls": intended IPC-weighted least-squares AFT objective; see the release defects below before relying on it.
• "squared": squared-error time-oriented objective.

Time-oriented losses do not make predict() a Cox risk score and do not provide the same baseline survival-function interface. Both non-Cox losses return exponentiated predictions on the original-time scale. Confirm prediction direction before using concordance or dynamic AUC; negate a predicted-time output only when that conversion is explicitly intended and reported.

Released 0.28.0 ipcwls has two native/source-confirmed limitations in both boosting classes: positive times below 1 become negative after the internal log transform and the censoring-weight estimator rejects them; and the gradient omits IPC weights even though the reported loss uses them. With fixed iterations, full subsampling, and default sample weights, changing event/censor labels while keeping observed times fixed produced identical predictions on the synthetic fixture. Do not treat a successful ipcwls fit as validated IPC-weighted training or change time units merely to hide the error. Prefer a verified alternative such as IPCRidge when an IPC-weighted AFT fit is required.

Current regularization controls:

• learning_rate
• n_estimators
• subsample
• dropout_rate
• tree depth/leaf controls
• ccp_alpha
• validation_fraction, n_iter_no_change, and tol
• a custom monitor passed to fit() for controlled early stopping

The old criterion parameter was removed in 0.28. Do not copy it from older examples.

Use only training data for internal early stopping. The final test set must not be the monitor or validation fraction.

Componentwise boosting

from sksurv.ensemble import ComponentwiseGradientBoostingSurvivalAnalysis

model = ComponentwiseGradientBoostingSurvivalAnalysis(
    loss="coxph",
    learning_rate=0.1,
    n_estimators=300,
    subsample=0.8,
    random_state=20260723,
)
model.fit(X_train_scaled, y_train)

At each iteration, a componentwise learner updates one encoded feature. The final model is linear and often sparse. Its coef_ includes the fitted intercept entry used by the implementation; align coefficients with transformed feature names carefully.

Iteration count is a selection parameter. Repeatedly checking test performance while increasing n_estimators leaks the test set. Tune it inside inner CV, and assess selected-feature stability across outer resamples.

Leakage-safe nested tuning

from sklearn.model_selection import GridSearchCV

inner_search = GridSearchCV(
    pipeline,
    {
        "model__min_samples_leaf": [3, 8, 16],
        "model__max_features": ["sqrt", 0.5, 1.0],
    },
    cv=inner_splits,
    error_score="raise",
    n_jobs=1,
)
inner_search.fit(X_outer_train, y_outer_train)
risk_outer = inner_search.predict(X_outer_valid)

Run this search inside each outer fold when reporting cross-validated tuned performance. Every outer score must use:

• outer-training preprocessing only;
• outer-training censoring distribution for IPCW metrics;
• an evaluation grid supported by that outer-training fold;
• outer-validation predictions never used in parameter selection.

After protocol assessment, tune on all development data and evaluate once on the reserved test set.

Probability prediction and calibration

Forests and Cox-loss boosting can produce survival probabilities. To use Brier metrics:

functions = fitted_pipeline.predict_survival_function(X_test)
survival_probability = np.vstack([fn(times) for fn in functions])

Then verify:

• shape is (n_test, n_times);
• values are within [0, 1];
• each row is non-increasing over time;
• times is strictly increasing and train-supported;
• censoring weights are learned from training outcomes.

A lower Brier score does not prove good calibration in every subgroup or horizon. Use horizon-specific calibration assessment on independent data. Do not claim clinical utility from concordance, AUC, or Brier score alone.

Practical selection questions

• Need coefficient-level PH interpretation? Start with a prespecified Cox model.
• Need nonlinear interactions and probability curves? Compare forest and Cox-loss boosting.
• Need sparse linear prediction? Compare Coxnet and componentwise boosting.
• Need time-oriented AFT prediction? Consider IPC-weighted losses and state the censoring assumptions.
• Need very large data? Benchmark memory and run time; kernel SVM and large survival forests can be expensive.

These are candidate-selection prompts, not performance guarantees.

Sources

Official sources checked 2026-10-01:

• Released survival losses and fit loops — IPCW caveats reproduced natively.

• Random survival forest user guide

• Gradient boosting user guide

• RandomSurvivalForest API

• ExtraSurvivalTrees API

• GradientBoostingSurvivalAnalysis API

• ComponentwiseGradientBoostingSurvivalAnalysis API

• 0.28 release notes

• 0.27 release notes