comprisk

A Python toolkit for competing-risks survival analysis: a scalable, scikit-learn-compatible competing-risks random survival forest plus the canonical regression / non-parametric methods — Fine-Gray, Aalen-Johansen CIF, cause-specific Cox — so applied researchers can drop the Python → R round-trip.

Status: alpha — API may change before v1.0. Renamed from crforest in 0.3.1 (pip install comprisk; from comprisk import CompetingRiskForest).

Highlights

• Four canonical CR methods, native Python — Fine-Gray (+ penalized), cause-specific Cox, Aalen-Johansen CIF, Gray's test — each validated to floating-point tolerance against cmprsk / crrp / survival.
• The only native-Python CR forest — composite & cause-specific CR log-rank splitting, AJ CIF, Nelson-Aalen CHF, Wolbers + Uno IPCW concordance, OOB Breiman VIMP, Ishwaran minimal depth, exact TreeSHAP.
• CR-aware evaluation — score_cr (IPCW time-dependent AUC/Brier + bootstrap CIs) and calibration_cr, replacing the CR-mode riskRegression::Score() block; concordance_index_ci gives a closed-form (no-bootstrap) CI for the Uno C-index and a paired model-comparison concordance_index_delta_ci.
• Fast — 10–22× vs randomForestSRC on real EHR, 16.6–544× vs scikit-survival (n = 5k → 50k), n = 10⁶ in 63 s — at matched C ≈ 0.85. Benchmarks →
• Reproducible — equivalence="rfsrc" reproduces rfSRC's per-tree mtry/nsplit RNG stream bit-for-bit. Methodology →

Install

pip install comprisk          # or:  uv add comprisk
pip install "comprisk[gpu]"   # CUDA 12 preview (faster only at low p today)

Python ≥ 3.10. Core deps: numpy, scipy, pandas, joblib, numba, scikit-learn.

Quickstart

from comprisk import CompetingRiskForest

# event: 0 = censored, k≥1 = cause-k event. Defaults: 100 trees, logrankCR, n_jobs=-1.
forest = CompetingRiskForest(n_estimators=200, random_state=42).fit(X, time, event)

cif  = forest.predict_cif(X[:5])          # (5, n_causes, n_times) — Aalen-Johansen
print(forest.oob_score(cause=1))          # honest out-of-bag C-index (no holdout split)
shap, base = forest.shap_values(X[:10])   # exact TreeSHAP (n, p, n_times, n_causes)

Prediction shapes, scoring, cross-validation, VIMP, minimal depth, GPU, and rfSRC migration — all with runnable code — are in the quickstart. Every estimator here is a real sklearn estimator — get_params / set_params / clone, and cross_val_score / Pipeline / GridSearchCV without a wrapper. Fit on a DataFrame and feature_names_in_ carries your column names through to VIMP output.

Like every survival library, these estimators cannot pass sklearn's full check_estimator suite: it supplies a plain numeric y, while competing-risks data needs both a time and an event per subject (use Surv.from_arrays). Everything outside that constraint — and sparse X, which is unsupported — is covered by tests/test_sklearn_compat.py.

Regression & non-parametric models

from comprisk import FineGrayRegression

fg = FineGrayRegression(cause=1, robust_se=True).fit(X, time=time, event=event)
print(fg.coef_, fg.se_)                    # log subdistribution-HRs

Worked code for every row is in examples/02_regression_models.ipynb.

comprisk vs alternatives

scikit-survival's CHF/survival outputs and scaling caveats are detailed in the benchmarks.

Benchmarks

Matched-pair, real EHR data (full tables + methodology in docs/benchmarks.md):

Both fit similarly well (C ≈ 0.85); the band tracks feature count. Also 16.6–544× vs scikit-survival (n = 5k → 50k) and n = 10⁶ in 63 s on a consumer i7.

Roadmap

comprisk is intentionally CR-focused — for non-CR survival (general Cox, AFT, deep-survival), use lifelines or scikit-survival.

• Shipped (v0.3–0.6): CR forest, Fine-Gray (+ penalized), cause-specific Cox, Aalen-Johansen CIF, Gray's test, score_cr / calibration_cr.
• v1.0 (planned): API freeze + JMLR MLOSS submission.
• v1.1 (planned): full GPU rewrite.

Documentation

📖 Full documentation site — searchable, autogenerated API reference.

• Quickstart — common tasks with runnable code
• API reference — full parameter lists
• Benchmarks — full tables, methodology, reproduction scripts
• Equivalence vs rfSRC — cross-library validation
• References — algorithmic provenance

Examples

Runnable notebooks in examples/ (rendered on GitHub; open in Colab to run):

• 01_forest_quickstart.ipynb — fit → CIF → OOB scoring → VIMP → minimal-depth selection
• 02_regression_models.ipynb — Fine-Gray, cause-specific Cox, AJ by group, Gray's test, penalized FG
• shap_explain.py — interactive marimo TreeSHAP app

Development

Requires uv.

uv venv && uv pip install -e ".[dev]"
uv run pre-commit install
uv run pytest && uv run ruff check .

License & citation

Apache-2.0 (LICENSE, NOTICE). If you use comprisk in your research, please cite the paper:

@article{yang_comprisk_2026,
  author        = {Yang, Sunny and Zhao, Weiyan and Zhao, Wanqi},
  title         = {{comprisk: A scikit-learn-compatible Python toolkit for competing-risks survival analysis}},
  year          = {2026},
  eprint        = {2607.09431},
  archivePrefix = {arXiv},
  primaryClass  = {stat.CO},
  url           = {https://arxiv.org/abs/2607.09431},
}

To cite a specific release, use the archived Zenodo DOI (concept-level 10.5281/zenodo.19876282, resolves to latest) or GitHub's "Cite this repository" button (CITATION.cff).