Breast Cancer Tumor Classification with Explainable AI

An end-to-end, reproducible comparison of logistic regression and random forest, with leakage-safe model selection and an untouched holdout.

Educational only — not a clinical tool

Built for learning on public data. It is not a medical device and must not be used for diagnosis.

Holdout accuracy: 96.49%
Holdout ROC-AUC: 0.9960
Malignant recall: 92.86%
CV fits: 105

Bar chart of global malignant-class feature importance on the holdout set (mean absolute SHAP value), led by worst texture. — FIG. 02.6 — Global SHAP feature importance · Figure from the project repository

01Focus

An end-to-end, reproducible comparison of logistic regression and random forest, with leakage-safe model selection and an untouched holdout.

02Data

UCI Wisconsin Diagnostic Breast Cancer dataset, loaded via scikit-learn: 569 samples, 30 features describing cell nuclei from digitized fine-needle aspirate images (212 malignant / 357 benign).

03Methods

Stratified 80/20 split (random_state 42); exploratory analysis on the training split only.
Pipelines with median imputation and scaling (logistic regression) or random forest, with preprocessing learned within each fold.
The same 5-fold stratified cross-validation for both models: 21 configurations, 105 CV fits.
A prespecified selection rule — highest mean training CV ROC-AUC, ties favor logistic regression — recorded before the holdout was touched.
Fixed 0.5 threshold, with no tuning on the holdout.
SHAP values for every holdout sample, checked for additivity (max error 3.6e-15).
One-command reproduction that regenerates the figures and README; continuous integration.

04Tools

Python
scikit-learn
SHAP
Streamlit
pytest
GitHub Actions

05Visualizations

SHAP beeswarm plot of per-sample feature contributions on the holdout set. — FIG. 02.6.1Figure from the project repository ·SHAP beeswarm on the holdout set.

Receiver operating characteristic curve for the selected model on the holdout set. — FIG. 02.6.2Figure from the project repository ·ROC curve on the holdout set.

06Findings

Logistic regression was selected.
Holdout (114 samples): 96.49% accuracy, 0.9960 ROC-AUC, 92.86% malignant recall.

07Limitations

Small historical sample, not a screening cohort.
No external or prospective validation.
One fixed holdout, with sampling uncertainty.
SHAP shows model associations, not biological causation.
Educational only — not a clinical tool.