CKD Classification — Minor Project

Machine-learning classification of Chronic Kidney Disease on the UCI CKD dataset (400 patients), evaluated with leak-free 5-fold stratified cross-validation.

Quick start (Google Colab — recommended for beginners)

1. Go to colab.research.google.com
2. File ▸ Upload notebook → upload notebooks/CKD_Classification_Colab.ipynb
3. Run the first cell and upload data/kidney_disease.csv when prompted
4. Runtime ▸ Run all

Colab already has pandas, numpy, scikit-learn, matplotlib and seaborn — nothing to install.

Quick start (local)

pip install pandas numpy scikit-learn matplotlib seaborn
python src/ckd_pipeline.py --experiment no_outlier       # Experiment 1 (primary)
python src/ckd_pipeline.py --experiment outlier_capped   # Experiment 2 (comparison)
PYTHONPATH=src python src/leakage_demo.py                # leakage check

python src/ckd_pipeline.py --mode normal                 # single train/test split only
python src/ckd_pipeline.py --mode kfold                  # 5-fold cross-validation only

Both classification modes run by default (--mode both):

• normal — one stratified 80/20 train/test split. One confusion matrix, one ROC curve and one set of metrics per model, on 80 patients the model has never seen.
• kfold — 5-fold stratified cross-validation: mean accuracy across the folds, 5 ROC curves per model (one per fold plus their mean), and a pooled out-of-fold confusion matrix covering all 400 patients.

Layout

data/kidney_disease.csv                   raw dataset (400 rows)
notebooks/CKD_Classification_Colab.ipynb  step-by-step teaching notebook (53 cells)
src/ckd_pipeline.py                       reproducible analysis script
src/leakage_demo.py                       measures leaky vs correct preprocessing
results/no_outlier/                       Experiment 1 tables (CSV) + figures (PNG)
  table_normal_classification.csv         holdout metrics per model
  table_model_summary.csv                 pooled out-of-fold + mean-accuracy columns
results/outlier_capped/                   Experiment 2 tables + figures
results/table_experiment_comparison.csv   side-by-side comparison
REPORT.md                                 methodology, results, discussion, conclusion

Holdout result (Experiment 1, single 80/20 split, n = 80)

Model	Accuracy	Sensitivity	Specificity	Precision	F1	AUC
Logistic Regression	0.9875	0.9800	1.0000	1.0000	0.9899	1.0000
Decision Tree	0.9750	0.9800	0.9667	0.9800	0.9800	0.9733
Random Forest	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
SVM (RBF)	0.9750	0.9600	1.0000	1.0000	0.9796	1.0000

On 80 test patients one misclassification is worth 0.0125 accuracy, so the whole spread above is two patients. The cross-validated numbers below use all 400 and are the ones to quote.

Headline result (Experiment 1, pooled out-of-fold, n = 400)

Model	Accuracy	Sensitivity	Specificity	Precision	F1	AUC
Logistic Regression	0.9925	0.9920	0.9933	0.9960	0.9940	0.9998
Decision Tree	0.9800	0.9880	0.9667	0.9802	0.9841	0.9773
Random Forest	0.9900	0.9960	0.9800	0.9881	0.9920	0.9997
SVM (RBF)	0.9950	0.9920	1.0000	1.0000	0.9960	1.0000

Positive class = CKD. Random Forest missed only 1 of 250 CKD patients (FN = 1) and is recommended for screening; SVM leads on accuracy and specificity.

Caveat: the top predictors are the defining diagnostic criteria of CKD, so this benchmark is near-separable. These scores would not transfer to raw hospital data. See REPORT.md for the full discussion.

How data leakage is prevented

Imputation, scaling, one-hot encoding and (in Experiment 2) IQR fences all live inside a sklearn.pipeline.Pipeline, which is refitted from scratch inside every training fold. Held-out folds are only .transform()-ed, never .fit()-ted on.