CKD / GFR Regression Analysis

Language: English | Русский

---

English

Project description

Machine learning project: predict GFR (kidney function) from clinical tabular features.

Methodology fix (important)

CKD_Status is not used in the final model.
CKD labels are typically derived from GFR stages, so using CKD status to predict GFR would be circular / leakage-like.

High R² is still expected because GFR is strongly linked to Creatinine (correlation ≈ -0.96).

Results (test set, without CKD_Status)

Model	Test R²	MAE	RMSE	CV R²
Gradient Boosting	0.9983	0.818	1.336	0.9978
Linear Regression	0.9909	1.754	3.095	0.9899

Leakage check (Linear Regression with CKD_Status, not used finally): R² ≈ 0.9931

Pipeline

1. EDA + correlation checks
2. Drop CKD_Status from features
3. Train/test split
4. Encode Medication on train only
5. Compare Linear Regression vs Gradient Boosting (CV + test)
6. Residual diagnostics + feature importance

How to run

pip install -r requirements.txt
jupyter notebook notebooks/ckd_regression_analysis.ipynb

Repository structure

ckd-regression-analysis/
├── notebooks/ckd_regression_analysis.ipynb
├── data/kidney_dataset.csv
├── data/DATA.md
├── images/
│   ├── residuals_gb.png
│   ├── feature_importance.png
│   └── creatinine_vs_gfr.png
├── README.md
├── README.ru.md
├── requirements.txt
├── LICENSE
└── .gitignore

Example plots

Limitations

• Educational / portfolio project, not a clinical tool
• Near-formula relationship Creatinine→GFR can inflate R²
• Dataset provenance should be treated carefully (see data/DATA.md)

Author

AnnaKazarian13