clinikit 0.1.0

A lightweight, sklearn-compatible Python toolkit for tabular machine learning.

clinikit

Prepared by Berat Kaan SEVEN

A lightweight, sklearn-compatible Python toolkit for tabular machine learning. clinikit bundles 14 hybrid classifiers, 5 experiment protocols, calibration utilities, label-noise diagnostics, fairness audits, and structured HTML reports behind a single drop-in package.

Research and development use only. This is an integration toolkit, not a regulated product and not a research paper of original methods. See CITATIONS.md for source-method references.

---

Why clinikit

clinikit is a complement to existing libraries, not a competitor.

Library	Focus	Why clinikit is different
scikit-learn	General-purpose ML	Adds curated experiment protocols, audit utilities, and structured reporting
Cleanlab	Label noise only	Integrates Cleanlab plus neighborhood conflict and LOO into one diagnostics module
MAPIE	Conformal prediction only	Includes selective classification as one of 14 bundled models
Fairlearn / AIF360	Fairness only	The audit module bundles fairness, leakage, and documentation helpers
AutoGluon	AutoML	Library-first; thin AutoML wrappers exist but no auto-magic by default
PyHealth	Deep learning for sequence / multimodal	Tabular-only, classical ML focused, lightweight

---

Installation

pip install clinikit

Optional dependency groups:

pip install "clinikit[diagnostics]"   # Cleanlab-based label-noise tools
pip install "clinikit[explain]"       # SHAP and LIME wrappers
pip install "clinikit[automl]"        # TabPFN, FLAML, AutoGluon wrappers
pip install "clinikit[synthetic]"     # CTGAN / TVAE wrappers
pip install "clinikit[conformal]"     # MAPIE conformal prediction
pip install "clinikit[all]"           # Everything

Supported Python versions: 3.10, 3.11, 3.12, 3.13.

---

Quickstart

from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

from clinikit.datasets import load_pima
from clinikit.metrics import sensitivity, specificity
from clinikit.models import RuleAugmentedClassifier

X, y = load_pima(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)

model = RuleAugmentedClassifier(base_estimator=LogisticRegression(max_iter=1000))
model.fit(X_train, y_train)
y_pred = model.predict(X_test)

print("Sensitivity:", sensitivity(y_test, y_pred))
print("Specificity:", specificity(y_test, y_pred))

For a complete walkthrough, see examples/quickstart.ipynb or open it in Colab.

---

What is in the box

14 hybrid classifiers (clinikit.models)

All sklearn-compatible, all pass sklearn.utils.estimator_checks.check_estimator.

• RuleAugmentedClassifier
• BoundaryRefineClassifier
• SubgroupThresholdClassifier
• ErrorAwareCalibrator
• MonotonicBooster
• HardSampleWeightedEnsemble
• ClassConditionalImputer
• CrossDistributionDistiller
• SelectiveClassifier
• InstanceAdaptiveThreshold
• DialecticalEnsemble
• LatentSubtypeRouter
• IterativeLabelRefiner
• DualViewCoTrainer

Supporting modules

• preprocessing — imputers, scalers, outlier flags, missing indicators
• metrics — sensitivity, specificity, NPV, PPV, F2, MCC, Brier, ECE
• curves — ROC, PR, calibration, Decision Curve Analysis
• protocols — 5 experiment protocols (Defensible, MaxScore, OriginalOnly, Deployment, Audit)
• leaderboard — experiment tracking CSV with 38 columns
• report — HTML structured report generator (Jinja2 templates)
• audit — leakage detection, subgroup fairness, documentation checks
• governance — audit-trail manifest templates (documentation only)
• reproducibility — manifest files (data hash + config + library versions)
• datasets — UCI benchmarks (PIMA, Wisconsin, UCI Heart, Frankfurt)
• cli — Typer-based CLI: train, benchmark, audit, validate, report
• thresholds, calibration, statistics, diagnostics, cost_sensitive, monitor, modelcard, cross_val, explainability, automl, external_val, time_split, active_learning, synthetic

---

Command-line interface

clinikit train      --config config.yaml
clinikit benchmark  --dataset pima --models all
clinikit audit      --data data.csv --report audit.html
clinikit validate   --model model.joblib --data data.csv
clinikit report     --leaderboard runs.csv --out report.html

---

Project notes

clinikit is an integration toolkit. The methods it bundles are adaptations of techniques published in the academic literature; see CITATIONS.md for source-method references. It is not a research paper of original methods, and it is not a regulated product. Research and development use only.

---

Contributing

Contributions are welcome. Please read CONTRIBUTING.md for the development workflow, coding standards, and pull-request process. By participating, you agree to abide by the Code of Conduct.

---

Citation

If you use clinikit in academic work, please cite it via the CITATION.cff file, or use:

@software{clinikit,
  author  = {SEVEN, Berat Kaan},
  title   = {clinikit: a tabular machine-learning toolkit},
  year    = {2026},
  url     = {https://github.com/clinikit/clinikit},
  version = {0.1.0}
}

---

License

Distributed under the MIT License. See LICENSE for the full text.