临床研究生存分析｜个人思维图谱小站

正文

references/cox-models.md

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

Cox proportional hazards model

For covariates (x),

[ h(t \mid x) = h_0(t)\exp(x^\top\beta). ]

CoxPHSurvivalAnalysis estimates coefficients by partial likelihood. The model assumes covariate effects multiply the hazard by a time-constant factor. A fitted coefficient is a log hazard ratio only under the model, coding, scale, and PH assumptions.

from sksurv.linear_model import CoxPHSurvivalAnalysis

model = CoxPHSurvivalAnalysis(
    alpha=0.1,
    ties="efron",
    n_iter=100,
    tol=1e-9,
)
model.fit(X_train, y_train)
risk = model.predict(X_test)
survival = model.predict_survival_function(X_test)
hazard = model.predict_cumulative_hazard_function(X_test)

Current key parameters:

alpha: non-negative L2/ridge penalty. It may be a scalar or feature-specific vector where documented. alpha=0 is unpenalized.
ties: "breslow" (default) or "efron".
n_iter, tol, verbose: Newton-Raphson controls.

predict() returns the linear predictor (x^\top\hat\beta); higher means higher event risk. Absolute survival probabilities come from the fitted baseline survival, not from transforming a risk score by itself.

Stability and interpretation

Encode and scale inside a training-fitted pipeline.
Use ridge shrinkage for unstable or correlated designs; a successful numerical fit does not establish inferential validity.
Check coefficient sensitivity to coding, scaling, missingness, influential rows, and regularization.
Exponentiating a coefficient gives a model-based hazard ratio for one unit of its encoded feature, holding other modeled features fixed.
A hazard ratio is not a risk ratio, probability difference, causal effect, or clinical utility measure.

scikit-survival does not provide a complete PH-diagnostics workflow. Assess the PH assumption using residual/graphical/domain methods appropriate to the study. If it fails, consider time interactions, stratification in a method that supports it, a time-varying model, an AFT model, or a flexible prediction model. Do not merely switch models and retain Cox coefficient interpretation.

Penalized Cox path with Coxnet

CoxnetSurvivalAnalysis implements a Cox elastic-net path:

[ \text{penalty} = \alpha\left(\rho|\beta|_1 + \frac{1-\rho}{2}|\beta|_2^2\right), ]

where l1_ratio is (\rho).

from sksurv.linear_model import CoxnetSurvivalAnalysis

model = CoxnetSurvivalAnalysis(
    n_alphas=100,
    alpha_min_ratio="auto",
    l1_ratio=0.9,
    fit_baseline_model=True,
)
model.fit(X_train_scaled, y_train)

Current details:

l1_ratio must be in (0, 1]; 1.0 is LASSO and values below 1 mix L1/L2. Exact pure ridge is handled by CoxPHSurvivalAnalysis(alpha=...), not by setting l1_ratio=0.
alphas=None estimates a decreasing path; explicit alphas selects the path.
alpha_min_ratio controls the smallest/largest path ratio. It is not the L1/L2 mixing parameter.
penalty_factor can vary penalties by feature; zero leaves a feature unpenalized.
normalize=False is current. Prefer an explicit StandardScaler pipeline so fold behavior and feature scaling are visible.
coef_ has shape (n_features, n_alphas). There is no current coef_path_ attribute.
predict(X, alpha=...) uses the selected path point (or interpolation).
predict_survival_function() and predict_cumulative_hazard_function() require fit_baseline_model=True. An explicit alpha for either probability method must match a fitted alphas_ value; risk-score interpolation does not provide an interpolated baseline survival model.

Leakage-safe alpha selection

Do not estimate the alpha path on all rows and then claim nested-CV performance. For a final held-out evaluation:

split train/test;
define an alpha grid from subject-matter scale or from training data only;
fit scaler and Coxnet inside each inner fold;
tune alpha on inner validation folds;
evaluate the selected procedure in an outer fold or untouched test set.

from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

pipeline = make_pipeline(
    StandardScaler(),
    CoxnetSurvivalAnalysis(l1_ratio=0.9, fit_baseline_model=True),
)
search = GridSearchCV(
    pipeline,
    {
        "coxnetsurvivalanalysis__alphas": [
            [0.01],
            [0.05],
            [0.2],
        ]
    },
    cv=inner_splits,
    error_score="raise",
)
search.fit(X_outer_train, y_outer_train)

When using censoring-aware scorer wrappers, wrap the entire pipeline and prefix the parameter again:

from sksurv.metrics import as_concordance_index_ipcw_scorer

wrapped = as_concordance_index_ipcw_scorer(pipeline, tau=tau)
search = GridSearchCV(
    wrapped,
    {
        "estimator__coxnetsurvivalanalysis__alphas": [
            [0.01],
            [0.05],
            [0.2],
        ]
    },
    cv=inner_splits,
)

IPCRidge .score() handles its time direction, but metric scorer wrappers call raw .predict() and do not automatically negate time-oriented predictions. Use an explicit risk adapter if applying a discrimination wrapper to an AFT estimator.

The wrapper is the estimator passed to GridSearchCV; it is not a zero-argument scoring callable. Its fit fold supplies the censoring distribution. Time support still has to be valid in every score fold.

Feature selection is uncertain

Non-zero coefficients at one alpha do not prove that a feature is biologically, causally, or clinically important. Report:

the exact preprocessing and penalty grid;
nested-CV or holdout protocol;
coefficient stability across resamples;
correlated alternatives and selection frequency;
the selected alpha and l1_ratio;
calibration and discrimination on independent data.

IPC ridge AFT model

IPCRidge is an inverse-probability-of-censoring weighted ridge regression model for a log-time/AFT objective:

from sksurv.linear_model import IPCRidge

model = IPCRidge(alpha=1.0)
model.fit(X_train_scaled, y_train)
predicted_time = model.predict(X_test_scaled)

The fitted linear predictor models log time, but predict() exponentiates it and returns original-time values; do not exponentiate it again. This output is time-oriented: larger predicted values imply longer predicted survival time, unlike higher-is-riskier Cox scores. Do not pass it unchanged to metrics expecting higher event risk. If a discrimination analysis requires a risk direction, use the negative prediction and state that transformation.

IPCW estimation relies on censoring assumptions and support. High censoring is not by itself a license to prefer the model; inspect weight stability and whether the training censoring distribution is positive over the target range.

Time-dependent prediction

For Cox PH:

[ S(t \mid x) = S_0(t)^{\exp(x^\top\beta)}. ]

Evaluate returned step functions on a shared, train-supported grid:

import numpy as np

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

The result is (n_test, n_times) and is suitable for Brier metrics when times also satisfies the metric's test/training support constraints. Extrapolation beyond learned event-time support is not justified.

Calibration and claims

Risk ranking can remain similar after a monotone transformation while probability calibration changes. Therefore:

report C-index or dynamic AUC as discrimination;
report Brier score as probability prediction error;
inspect horizon-specific calibration separately;
validate on data independent of fitting and tuning;
do not infer treatment effects from predictive Cox coefficients;
do not call a model clinically useful without decision-focused evaluation.

Metadata routing

Current estimators expose get_metadata_routing(). Coxnet also exposes set_predict_request(alpha=...) for passing its optional alpha prediction argument through a meta-estimator. This only matters when:

from sklearn import set_config

set_config(enable_metadata_routing=True)

and an enclosing meta-estimator is expected to route that metadata. Ordinary pipeline.fit(X, y) and direct pipeline.predict(X) do not require enabling it.

Sources

Official sources checked 2026-10-01:

临床研究生存分析

正文