Scikit Survival

Source: https://github.com/aipoch/medical-research-skills

When to Use

Use this skill when you need to:

1. Model time-to-event outcomes with censoring (right/left/interval censored observations).
2. Fit and interpret Cox Proportional Hazards models (including penalized Cox for high-dimensional data).
3. Train non-linear survival models such as Random Survival Forests or Gradient Boosting survival models.
4. Use Survival SVMs for margin-based survival prediction (linear or kernel).
5. Evaluate survival predictions with censoring-aware metrics (Uno/Harrell C-index, time-dependent AUC, Brier/Integrated Brier Score) and/or perform competing risks analysis.

Key Features

• Survival target construction via sksurv.util.Surv (arrays or DataFrame).
• Model families
  • Cox models: CoxPHSurvivalAnalysis, CoxnetSurvivalAnalysis
  • Ensembles: RandomSurvivalForest, GradientBoostingSurvivalAnalysis, ExtraSurvivalTrees
  • SVM-based: FastSurvivalSVM, FastKernelSurvivalSVM
• Non-parametric estimators: Kaplan–Meier and Nelson–Aalen.
• Competing risks: cumulative incidence estimation.
• scikit-learn compatibility: pipelines, cross-validation, and GridSearchCV with survival scorers.
• Evaluation utilities: IPCW-based metrics (e.g., Uno’s C-index) and calibration-aware scores (IBS).

Additional topic guides may exist under:

• references/cox-models.md
• references/ensemble-models.md
• references/svm-models.md
• references/data-handling.md
• references/evaluation-metrics.md
• references/competing-risks.md

Dependencies

• scikit-survival (recommended: >=0.22)
• scikit-learn (recommended: >=1.2)
• numpy (recommended: >=1.23)
• pandas (recommended: >=1.5)

Example Usage

A complete, runnable example using a scikit-survival built-in dataset, a scikit-learn pipeline, and Uno’s C-index (IPCW):

import numpy as np

from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler

from sksurv.datasets import load_breast_cancer
from sksurv.linear_model import CoxPHSurvivalAnalysis
from sksurv.metrics import concordance_index_ipcw, as_concordance_index_ipcw_scorer

# 1) Load data (X: features, y: structured array with fields like ('event', 'time'))
X, y = load_breast_cancer()

# 2) Split (keep y_train for IPCW-based metrics)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

# 3) Build a pipeline (scaling is important for many survival models)
pipe = Pipeline([
    ("scaler", StandardScaler()),
    ("model", CoxPHSurvivalAnalysis()),
])

# 4) Optional: hyperparameter tuning (CoxPH has few knobs; shown for workflow completeness)
# If your version exposes regularization parameters, tune them here.
param_grid = {
    # Example placeholder; remove if unsupported in your installed version:
    # "model__alpha": [0.0, 1e-4, 1e-3]
}

if param_grid:
    search = GridSearchCV(
        pipe,
        param_grid=param_grid,
        scoring=as_concordance_index_ipcw_scorer(),
        cv=5,
        n_jobs=-1,
    )
    search.fit(X_train, y_train)
    best = search.best_estimator_
else:
    best = pipe.fit(X_train, y_train)

# 5) Predict risk scores (higher typically means higher risk / shorter survival)
risk_scores = best.predict(X_test)

# 6) Evaluate with Uno's C-index (IPCW)
c_uno = concordance_index_ipcw(y_train, y_test, risk_scores)[0]
print(f"Uno's C-index (IPCW): {c_uno:.3f}")

Implementation Details

1) Survival Target Representation (Surv)

scikit-survival expects outcomes as a structured array with at least:

• an event indicator (boolean)
• a time value (float/int)

Common construction patterns:

from sksurv.util import Surv

y = Surv.from_arrays(event=event_array, time=time_array)
# or
y = Surv.from_dataframe("event", "time", df)

2) Model Selection Heuristics

• High-dimensional (p > n): prefer CoxnetSurvivalAnalysis (Elastic Net) for stability and feature selection.
• Interpretability required: prefer CoxPHSurvivalAnalysis (coefficients as log hazard ratios).
• Strong non-linearities / interactions: prefer RandomSurvivalForest or GradientBoostingSurvivalAnalysis.
• Kernelized decision boundaries: consider FastKernelSurvivalSVM (ensure scaling).

3) Preprocessing Requirements

• Scaling: strongly recommended for SVMs and often beneficial for penalized Cox models.
• Categoricals: encode (e.g., one-hot) before fitting most estimators.
• Data validation: ensure non-negative times; verify enough events relative to feature count.

4) Evaluation Under Censoring

• Harrell’s C-index (concordance_index_censored): common, but can be less robust with heavy censoring.
• Uno’s C-index (concordance_index_ipcw): uses inverse probability of censoring weights and requires y_train to estimate censoring distribution.

from sksurv.metrics import concordance_index_censored, concordance_index_ipcw

c_harrell = concordance_index_censored(y_test["event"], y_test["time"], risk_scores)[0]
c_uno = concordance_index_ipcw(y_train, y_test, risk_scores)[0]

5) Time-dependent Metrics (AUC, Brier/IBS)

• Time-dependent AUC evaluates discrimination at specific time horizons.
• Brier score / Integrated Brier Score (IBS) evaluates calibration + discrimination over time and requires survival probabilities/functions.

from sksurv.metrics import cumulative_dynamic_auc

times = np.array([365, 730, 1095])  # example horizons
auc, mean_auc = cumulative_dynamic_auc(y_train, y_test, risk_scores, times)

6) Competing Risks (Cumulative Incidence)

Use competing risks methods when multiple mutually exclusive event types exist and one event prevents the others.

from sksurv.nonparametric import cumulative_incidence_competing_risks

# y must encode event types appropriately for competing risks workflows
time_points, cif1, cif2 = cumulative_incidence_competing_risks(y)