tausurv

Survival analysis for Python: nonparametric estimators, regression models, tree ensembles and neural networks behind one interface, with support for competing risks and minimal dependencies.

Documentation: https://scai-bio.github.io/tausurv/

Installation

pip install tausurv

Requires Python 3.12 or newer. Optional extras:

extra	adds
tausurv[plot]	plotting with matplotlib
tausurv[boost]	SurvivalBoost, via scikit-learn
tausurv[tune]	hyperparameter search and nested cross-validation, via Optuna
tausurv[deepsurv]	the four DeepSurv benchmark tables stored as HDF5

The neural models in tausurv.nn need PyTorch, which is not installed automatically because the right build depends on your hardware. Install it first by following https://pytorch.org/get-started/locally/.

Example

import tausurv as ts

X, time, event = ts.datasets.load_dataset("pbc")
time = time / 365.25  # days -> years

km = ts.nonparametric.kaplan_meier(time, event)
km(5.0)  # 5-year survival, about 0.70

covariates = X.select("age", "bili", "albumin").to_numpy()
cox = ts.linear.CoxPH().fit(covariates, time, event)
cox.predict_survival_function(covariates[:3], [1.0, 5.0, 10.0])

What's included

	tausurv	scikit-survival	lifelines	pycox	hazardous
Nonparametric
Kaplan-Meier	✓	✓	✓
Nelson-Aalen	✓	✓	✓
Aalen-Johansen	✓	✓	✓
Left truncation (delayed entry)		✓	✓
Log-rank test		✓	✓
Parametric univariate fits			✓
Regression
Cox proportional hazards	✓	✓	✓
Penalized Cox (L1, L2, elastic net)		✓	✓
Time-varying covariates			✓
Proportional-hazards test			✓
AFT (Weibull, log-normal, log-logistic)	✓		✓
Aalen additive hazards			✓
Fine-Gray	✓
Trees and boosting
Survival tree	✓	✓
Random survival forest	✓	✓
Gradient boosting	✓	✓			✓
Competing-risks gradient boosting	✓				✓
Survival support vector machine		✓
Neural networks
DeepSurv	✓			✓
DeepHit, with competing risks	✓			✓
Logistic hazard	✓			✓
Cox-Time, MTLR, PC-Hazard				✓
Deep survival machines	✓
CopulaSurv (dependent censoring)	✓
HACSurv (dependent competing risks)	✓
Discrimination
Harrell's concordance	✓	✓	✓
Uno's concordance (IPCW)	✓	✓
Antolini's time-dependent concordance	✓			✓
Time-dependent AUC	✓	✓
Prediction error and calibration
Brier score (IPCW)	✓	✓		✓	✓
Integrated Brier score	✓	✓		✓	✓
Censored negative log-likelihood	✓
CRPS	✓
Calibration curve	✓		✓		✓
D-calibration	✓
Competing-risks metrics
Cause-specific concordance	✓				✓
Cause-specific Brier score	✓				✓
Cause-specific AUC	✓
Cause-specific calibration	✓				✓
Workflow
Survival-aware cross-validation and tuning	✓	partly
Nested cross-validation	✓
Save and load fitted models	partly	✓	✓	✓	✓
Bundled datasets	✓	✓	✓	✓
Survival curves with at-risk tables	✓		✓
Time-dependent SHAP plots	✓

Checked against scikit-survival 0.27.0, lifelines 0.30.3, pycox 0.3.0 and hazardous 0.2.0. The full comparison explains the partial cells.

Development

The repository is a uv workspace. Install PyTorch into it with the build for your hardware, then sync:

uv pip install torch --index-url https://download.pytorch.org/whl/cpu
uv sync --all-packages --all-groups
uv run pytest

License

MIT. Developed at Fraunhofer SCAI.