Chronic Kidney Disease (CKD) Prediction System

A machine learning project that predicts the presence of chronic kidney disease from routine clinical test results, using six ensemble classifiers and a Flask web app for real-time predictions.

Built for CSE6505 – Machine Learning (CCP Project), Lahore Garrison University.

---

Overview

This project trains and compares six ensemble machine learning models on the UCI Chronic Kidney Disease dataset, then serves the best-performing model through a simple web interface where a clinician can enter a patient's lab values and get an instant CKD risk prediction.

• Best model: Voting Classifier — 100% accuracy, 100% AUC-ROC on the held-out test set
• Dataset: 397 patients, 24 clinical attributes (11 numeric, 13 categorical)
• Pipeline: ARFF → CSV conversion, MICE/mode imputation, encoding, scaling, SVM-SMOTE balancing, 6-model ensemble training, full evaluation suite
• Deployment: Flask app with an interactive form-based UI

---

Project Structure

ckd_project/
├── app.py                          # Flask web application (prediction API + UI)
├── ckd_analysis.ipynb              # End-to-end notebook: EDA → preprocessing → training → evaluation
├── model_metrics.csv               # Summary table of all 6 models' performance
│
├── Chronic_Kidney_Disease/
│   ├── chronic_kidney_disease.arff         # Original UCI dataset
│   ├── chronic_kidney_disease_full.arff    # Full/cleaned ARFF source used by the notebook
│   └── chronic_kidney_disease.info.txt     # Official UCI attribute documentation
│
├── data/
│   └── ckd_raw.csv                 # Dataset converted from ARFF to CSV (397 rows × 25 cols)
│
├── models/                         # Trained models + preprocessing artifacts (joblib .pkl)
│   ├── preprocessors.pkl           # Scaler, imputers, label encoders, feature names
│   ├── Random_Forest.pkl
│   ├── Gradient_Boosting.pkl
│   ├── AdaBoost.pkl
│   ├── Voting_Classifier.pkl       # ← model used in production (app.py)
│   ├── Stacking_Classifier.pkl
│   └── Bagging_Classifier.pkl
│
├── outputs/                        # Generated plots and metrics from the notebook run
│   ├── 01_missing_class.png        # Missing-value heatmap + class distribution
│   ├── 02_numeric_distributions.png# Numeric feature distributions by class
│   ├── 03_correlation.png          # Correlation heatmap (numeric features)
│   ├── 04_categorical.png          # Categorical feature breakdowns
│   ├── 05_confusion_matrices.png   # Confusion matrix per model
│   ├── 06_model_comparison.png     # Bar chart comparing all metrics across models
│   ├── 07_roc_curves.png           # ROC curves for all models
│   ├── 08_feature_importance.png   # Feature importances from the best model
│   ├── 09_cross_validation.png     # 10-fold CV accuracy comparison
│   └── model_metrics.csv           # Same metrics table, written by the notebook
│
└── templates/
    └── index.html                  # Frontend form for the Flask app

---

Dataset

Source: UCI Machine Learning Repository – Chronic Kidney Disease Data Set, collected by Dr. P. Soundarapandian (Apollo Hospitals, Tamil Nadu, India) and donated by L. Jerlin Rubini, Alagappa University (2015).

• Instances: 400 patients originally (397 retained after cleaning) — 250 CKD / 150 not-CKD (approx., before cleaning)
• Attributes: 24 clinical features + class label

Code	Feature	Code	Feature
age	Age	pcv	Packed cell volume
bp	Blood pressure	wc	White blood cell count
sg	Specific gravity	rc	Red blood cell count
al	Albumin	htn	Hypertension
su	Sugar	dm	Diabetes mellitus
rbc	Red blood cells	cad	Coronary artery disease
pc	Pus cell	appet	Appetite
pcc	Pus cell clumps	pe	Pedal edema
ba	Bacteria	ane	Anemia
bgr	Blood glucose random	sc	Serum creatinine
bu	Blood urea	sod	Sodium
hemo	Hemoglobin	pot	Potassium

Full attribute descriptions are in Chronic_Kidney_Disease/chronic_kidney_disease.info.txt.

---

Methodology

The complete pipeline is implemented in ckd_analysis.ipynb:

1. Data loading — parses the raw .arff file and converts it to data/ckd_raw.csv.
2. Exploratory Data Analysis — missing value patterns, class balance, numeric/categorical distributions, correlation analysis (saved to outputs/01–04).
3. Preprocessing
   • Target encoding (ckd = 1, notckd = 0)
   • Categorical features label-encoded
   • Numeric features imputed via Iterative Imputer (MICE); categorical features imputed via most-frequent value
   • StandardScaler applied to all features
   • Train/test split (80/20, stratified) performed before resampling to avoid leakage
   • SVM-SMOTE applied to the training set only, to correct class imbalance
4. Model training — six ensemble classifiers trained on the resampled training set, evaluated on the untouched test set, with 10-fold stratified cross-validation for robustness:
   • Random Forest (200 trees, max_features='sqrt')
   • Gradient Boosting (200 estimators, depth 4, lr 0.05)
   • AdaBoost (depth-2 stumps, 100 estimators)
   • Voting Classifier (soft voting: RF + GB + Logistic Regression)
   • Stacking Classifier (RF + GB + KNN → Logistic Regression meta-learner)
   • Bagging Classifier (100 decision trees)
5. Evaluation — accuracy, precision, recall, F1, AUC-ROC, and cross-validated accuracy for every model, plus confusion matrices, ROC curves, and feature importance plots (outputs/05–09).
6. Artifact export — all trained models and the fitted preprocessing objects are saved to models/ via joblib.

---

Results

Model	Accuracy	Precision	Recall	F1-Score	AUC-ROC	CV Mean ± Std
Random Forest	98.75%	100.00%	98.00%	98.99%	100.00%	99.00% ± 1.22%
Gradient Boosting	95.00%	96.00%	96.00%	96.00%	99.73%	97.00% ± 2.45%
AdaBoost	100.00%	100.00%	100.00%	100.00%	100.00%	99.50% ± 1.00%
Voting Classifier	100.00%	100.00%	100.00%	100.00%	100.00%	99.75% ± 0.75%
Stacking Classifier	97.50%	96.15%	100.00%	98.04%	100.00%	99.50% ± 1.00%
Bagging Classifier	96.25%	96.08%	98.00%	97.03%	99.67%	96.75% ± 2.51%

The Voting Classifier was selected as the production model (highest accuracy with the lowest cross-validation variance) and is the model loaded by app.py.

Note: near-perfect scores on a small (397-row), single-source dataset are expected and should be interpreted cautiously — see Limitations below.

---

Web Application

app.py is a lightweight Flask app that:

• Serves a clinical data-entry form (templates/index.html)
• Accepts patient lab values via a /predict POST endpoint
• Applies the saved preprocessors (imputers + scaler) and the Voting Classifier model
• Returns a prediction (CKD Detected / No CKD Detected), CKD probability, and a risk level (Low / Medium / High)

Running it locally

pip install flask numpy joblib scikit-learn
python app.py

Then open http://localhost:5000 in your browser.

Note: app.py currently loads the model from a hardcoded Windows path:

MODEL_DIR = r'C:\Users\SK COMPUTERS\Documents\ckd_project\models'

Update this to a relative path (e.g. MODEL_DIR = os.path.join(os.path.dirname(__file__), 'models')) before running on another machine.

---

Requirements

flask
numpy
pandas
scikit-learn
imbalanced-learn
matplotlib
seaborn
joblib
scipy

---

Limitations & Disclaimer

• The dataset is small (397 patients) and from a single hospital source, which inflates apparent model performance and limits generalizability to other populations.
• This tool is built for an academic course project and is not a certified medical device. It should not be used for actual clinical diagnosis or treatment decisions without validation by qualified medical professionals.

---

Acknowledgements

• Dataset: Dr. P. Soundarapandian (Apollo Hospitals), L. Jerlin Rubini & Dr. P. Eswaran (Alagappa University) — via the UCI Machine Learning Repository.