pysurvex

Model-agnostic explainability for survival analysis models in Python.

pysurvex is a Python adaptation of the R package survex. Its goal is to provide a unified, model-agnostic interface for explaining the predictions of survival analysis models (Cox PH, Random Survival Forests, DeepSurv, etc.) using time-dependent variants of popular interpretability methods.

Supported models

Auto-detected (no configuration needed):

• scikit-survival CoxPHSurvivalAnalysis
• scikit-survival RandomSurvivalForest
• pycox DeepSurv (pycox.models.CoxPH)

Any other survival model is supported by supplying a manual predict_survival_function override (see below).

Roadmap

Currently implemented xAI methods:

• Partial Dependence Profiles (PDP) — via model_profile(..., method="pdp")
• Accumulated Local Effects (ALE) — via model_profile(..., method="ale")
• Individual Conditional Expectation (ICE) — via predict_profile(...)

Methods to be implemented

• SurvLIME
• SurvSHAP

Installation

For local development:

git clone https://github.com/vedvatsal3/pysurvex.git
cd pysurvex
pip install -e ".[dev]"

Runtime dependencies: numpy, pandas, scikit-survival, scikit-learn. matplotlib is additionally required only when you call a result object's .plot() method — the numeric layer (computing PDP / ICE / ALE matrices) needs only numpy, pandas, and scikit-survival.

Supplying a custom prediction interface

For models that are not auto-detected, pass a predict_survival_function with the signature (model, X, times) -> np.ndarray of shape (n_obs, n_times):

def my_predict(model, X, times):
    # e.g. wrap pycox, lifelines, torchsurv, ...
    return model.survival_matrix(X, times)

explainer = SurvivalExplainer(
    model=my_model,
    data=X,
    y=y,
    predict_survival_function=my_predict,
)

References

• R package: https://modeloriented.github.io/survex/
• Paper: Spytek et al., survex: an R package for explaining machine learning survival models, Bioinformatics, 2023.

License

MIT -- see LICENSE.