crforest 0.1.1

Competing risks random survival forests for Python. 4.5–6× faster than randomForestSRC, scales to n=10⁶ in minutes, scikit-learn-compatible.

crforest

Competing-risks random survival forests for Python. 4.5–6× faster than randomForestSRC, scales to n = 10⁶ in minutes, scikit-learn-compatible. Designed to replace the Python → R workflow split that applied researchers currently endure for competing-risks survival analysis.

Status: pre-alpha (v0.1). API and internals may change before v1.0.

Highlights

• The only competing-risks Random Survival Forest in Python. Three-state fit and predict, Aalen-Johansen CIF, Nelson-Aalen CHF, cause-specific Harrell + Uno IPCW C-indices, OOB Breiman permutation VIMP — out of the box.
• 4.5–6× faster than randomForestSRC at matched accuracy (HF Harrell C tied at 0.864, real CHF n ≈ 75k); ~58× faster than rfSRC built without OpenMP (default R-on-macOS).
• Order-of-magnitude faster than scikit-survival (5.4× at n = 5k, 192× at n = 50k), without disabling CIF/CHF outputs.
• Bit-identical to randomForestSRC with equivalence="rfsrc" — reproduces the per-tree mtry/nsplit RNG stream for paper-grade reproducibility, sensitivity checks, and rfSRC-baseline migrations.

crforest vs alternatives

	crforest	randomForestSRC	scikit-survival
Language	Python	R	Python
Native competing risks	✓	✓	✗ (single-event only)
Aalen–Johansen CIF output	✓	✓	n/a
Cumulative hazard at scale	✓	✓	✗¹
OOB permutation VIMP	✓	✓	✗
Bit-identical reproducibility mode	✓ (equivalence="rfsrc")	—	n/a
Scales to n = 10⁶	✓	OOM at n ≳ 500 000	✗¹ / OOM²
Default parallelism	✓ (n_jobs=-1)	OpenMP (build-dependent; macOS Apple clang lacks it)	✓
GPU preview	✓ (CUDA 12)	✗	✗

¹ sksurv RandomSurvivalForest(low_memory=True) is the only mode that scales beyond ~10k samples, but it disables predict_cumulative_hazard_function and predict_survival_function (raises NotImplementedError). ² sksurv low_memory=False exposes CHF / survival outputs but stores per-leaf full CHF arrays; peak RSS reaches 16.8 GB at n = 5k on synthetic, OOMs (> 21.5 GB) at n = 10k on a 24 GB host.

Install

pip install crforest          # or:  uv add crforest
pip install "crforest[gpu]"   # or:  uv add 'crforest[gpu]'

Requires Python ≥ 3.10. Core dependencies: numpy, scipy, pandas, joblib, numba, scikit-learn. GPU extra adds cupy + CUDA 12 runtime libs (preview; faster only at low feature count today, full rewrite scheduled for v1.1).

Quickstart

import numpy as np
from crforest import CompetingRiskForest

# Toy competing-risks data: 500 subjects, 6 features, 2 causes (+ censoring).
rng = np.random.default_rng(42)
n = 500
X = rng.normal(size=(n, 6))
time = rng.exponential(2.0, size=n) + 0.1
event = rng.choice([0, 1, 2], size=n, p=[0.4, 0.4, 0.2])  # 0 = censored

# Fit. Defaults: n_estimators=100, max_features="sqrt", logrankCR, n_jobs=-1.
forest = CompetingRiskForest(n_estimators=100, random_state=42).fit(X, time, event)

# Per-subject risk score for cause 1 (suitable for Wolbers C-index).
risk = forest.predict_risk(X[:5], cause=1)

# Aalen-Johansen cumulative incidence over the forest's chosen time grid.
cif = forest.predict_cif(X[:5])                       # (5, n_causes, n_times)
cif_at = forest.predict_cif(X[:5], times=[1.0, 2.0, 5.0])

# Cause-specific Wolbers concordance.
print("C-index, cause 1:", forest.score(X, time, event, cause=1))

# OOB permutation VIMP, scored with Uno IPCW.
vimp = forest.compute_importance(random_state=42)
print(vimp.sort_values("composite_vimp", ascending=False).head())

See docs/quickstart.md for the full walkthrough — data format, prediction shapes, cross-validation, GPU, and migrating from rfSRC.

scikit-learn drop-in

CompetingRiskForest is a real sklearn estimator: BaseEstimator subclass, clone()-friendly, picklable, and fit(X, y) / score(X, y) accept the scikit-survival-style structured y so cross_val_score, KFold, and Pipeline work without a wrapper.

from sklearn.model_selection import KFold, cross_val_score
from crforest import CompetingRiskForest, Surv

y = Surv.from_arrays(event=event, time=time)
forest = CompetingRiskForest(n_estimators=100, random_state=42, n_jobs=-1)

cv = KFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(forest, X, y, cv=cv, n_jobs=-1)
print(f"5-fold C-index, cause 1: {scores.mean():.3f} ± {scores.std():.3f}")

The legacy three-argument form forest.fit(X, time, event) keeps working, and predict(X) is an alias for predict_risk(X, cause=1) so the estimator slots straight into Pipeline / cross_val_predict. For cause-k risk or for CIF / CHF curves use the explicit methods.

Benchmarks

vs randomForestSRC, paired same machine — i7-14700K, 28 threads, rfSRC 3.6.2 built with full OpenMP; real CHF cohort, HF / death competing risks; n = 75 000, p = 58, ntree = 100, 3 seeds (mean ± std):

	crforest (n_jobs=-1)	rfSRC OMP-on (rf.cores=28)	rfSRC OMP-off (rf.cores=1)
Wall time	17.75 ± 0.36 s	80.81 ± 0.68 s	1026.3 ± 6.6 s (~17 min)
Peak RSS	4.36 GB	22.70 GB	17.80 GB
HF Harrell C-index	0.8642	0.8645	0.8645
Speedup vs crforest	—	4.55×	57.8×

The OMP-off column is the configuration most R-on-macOS users hit out of the box — rfSRC's OpenMP requires rebuilding R against Homebrew gcc/clang with OpenMP support, which most Mac R installs lack. rfSRC OMP-on and OMP-off produce algorithmically identical output at the same seed (we verified per-seed err.rate is bit-identical), so the OMP-off C-index is taken from the OMP-on cell at the same seed. Reproducible via validation/comparisons/n75k_path_b.py.

Apples-to-apples vs rfSRC's best ntime config on the same workload: 6.13×. Speedup ratio is stable across ntree ∈ {100, 500, 1000}.

vs scikit-survival, paired same machine — i7-14700K, 28 threads, synthetic 2-cause Weibull DGP, p = 58, ntree = 100, both libraries at their best config (n_jobs=-1; sksurv low_memory=True):

n	sksurv low_memory=True	crforest	speedup
5 000	18.2 s / 0.22 GB peak RSS	3.4 s / 1.83 GB	5.4×
10 000	85.0 s / 0.26 GB	5.6 s / 3.06 GB	15.1×
25 000	609.7 s / 0.37 GB	10.4 s / 4.82 GB	58.4×
50 000	2 935.3 s (49 min) / 0.55 GB	15.3 s / 6.80 GB	191.6×

The wall-time gap widens with n (sksurv RSF wall scales ≈ n^2.2 with default min_samples_leaf=3; crforest histogram split kernel ≈ n^0.6). At every paired point crforest also provides Aalen-Johansen CIF, Nelson-Aalen CHF, and risk scores; sksurv low_memory=True provides only predict() risk scores — predict_cumulative_hazard_function and predict_survival_function raise NotImplementedError. When both are scored on the single-event-collapsed truth sksurv is fit on, holdout Harrell C-index is matched within ±0.01 across n; crforest additionally surfaces a cause-specific Wolbers concordance of 0.69–0.72 for cause 1 (HF) that sksurv has no way to compute on competing-risk data. Sksurv's other mode (low_memory=False) restores CHF/survival outputs but OOMs: at n = 5k it already peaks at 16.8 GB RSS; at n = 10k it exceeds a 21.5 GB cap on a 24 GB host. Numbers reproducible via validation/comparisons/sksurv_oom.py.

Scaling (one-sided beyond the paired ranges). crforest exhibits sub-linear wall growth in n with the histogram split kernel:

Workload (default config, ntree = 100)	crforest CPU wall	rfSRC	sksurv
n = 75 000 (real CHF, paired)	17.75 s	80.81 s	OOM (low_memory=False) / extrapolated ~2.4 hr (low_memory=True)
n = 1 000 000 (UKB-scale feasibility)	122 s	not measured (rfSRC peak RSS at n = 75 000 already 22.7 GB OMP-on / 17.8 GB OMP-off; extrapolation puts n = 500 000 past 80 GB)	not feasible at full output capability

We do not publish a paired number above n ≈ 100 000 because both R and Python alternatives are impractical there, not because the comparison gets unfavourable.

Even when forced to n_jobs=1, crforest stays fast: the histogram split kernel is numba-compiled, so a single-thread fit uses the same hot loop as the parallel one (just without the joblib fan-out). That matters for environments where worker pools are forbidden — Jupyter notebooks under certain spawn configs, locked-down HPC schedulers, nested-parallel pipelines, etc.

The compact leaf storage (per-cause integer event/at-risk counts only; CIF/CHF tables materialise lazily on first predict) keeps pickle size proportional to the cohort: n = 100k, ntree = 100 pickles to ~5.0 GB.

API (one-line summary)

CompetingRiskForest(...)	the estimator; full parameter list in forest.py
.fit(X, time, event) or .fit(X, y)	legacy form takes time (n,) float and event (n,) int (0=censored, 1..K=cause); sklearn form takes structured y from Surv.from_arrays
.predict(X)	sklearn alias for predict_risk(X, cause=1); shape (n_samples,)
.predict_cif(X, times=None)	shape (n_samples, n_causes, n_times)
.predict_chf(X, times=None)	shape (n_samples, n_causes, n_times)
.predict_risk(X, cause=1, kind="integrated_chf")	shape (n_samples,)
.score(X, time, event, cause=1) or .score(X, y, cause=1)	Wolbers cause-specific concordance
.compute_importance(...)	per-cause + composite permutation VIMP DataFrame
Surv.from_arrays(event, time)	structured y for sklearn fit(X, y) / cross_val_score
concordance_index_cr(event, time, estimate, cause=1)	top-level metric for any risk score

After .fit(): forest.n_causes_, forest.unique_times_, forest.time_grid_, forest.n_features_in_, forest.feature_importances_, forest.trees_, forest.inbag_ (when equivalence="rfsrc").

Two splitrules are available: logrankCR (composite competing-risks log-rank, default) and logrank (cause-specific).

Documentation

• Quickstart — common tasks with runnable code
• PRD — what crforest aims to be at v1.0
• Equivalence vs rfSRC — cross-library validation methodology
• References — algorithmic provenance (Park-Miller, Bays-Durham, Wolbers 2009, Uno 2011, Cole & Hernán 2008, Breiman 2001, Ishwaran 2008/2014, etc.)

Development

Requires uv.

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

License

Apache-2.0. See LICENSE and NOTICE.

Citation

A paper describing crforest is in preparation. Until it is out, please cite the project URL and version. Algorithmic references (Park-Miller, Bays-Durham, Wolbers, Uno, Cole & Hernán, Kaplan-Meier, Breiman, Ishwaran) are listed in docs/REFERENCES.md.