Apppublic

workface/coronary-artery-disease

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

App README

Coronary Artery Disease Risk Estimator

Feature-selection-driven binary classifier for coronary artery disease, trained on the UCI Cleveland heart disease dataset and deployed as a Streamlit app.

![Python](https://www.python.org/) ![License](LICENSE) ![Live Demo](https://huggingface.co/spaces/workface/coronary-artery-disease)

Live demo: huggingface.co/spaces/workface/coronary-artery-disease

[image]

[!WARNING] Educational project — not medical advice. This model is trained on 303 records from a 1988 UCI dataset and has not been clinically validated. Do not use for diagnosis, treatment, or any medical decision.

Problem

Coronary artery disease (CAD) is typically confirmed by angiography — an invasive, expensive procedure. This project asks: given routinely-collected patient data (demographics, stress-test results, ECG findings), how accurately can we predict CAD before angiography, and which features actually matter? The answer informs which non-invasive measurements clinicians should prioritise.

Dataset

—Source: UCI Cleveland heart disease dataset
—Size: 303 patients, 14 columns
—Target: ca_disease (0 = no disease, 1–4 = increasing severity → binarised to 0/1)
—Class balance: roughly 54% / 46% — no rebalancing required.
—Missing values: 6 total (? sentinels in num_major_vessels and thallium_scint) → median imputation.

Methodology

1.Cleaning — coerce ? to NaN, median-impute the 6 missing values, remap thallium_scint codes {3:0, 7:1, 6:2} so they are ordered by severity, binarise the target.
2.EDA — distributions, correlations, violin plots for continuous features, bar charts for categoricals.
3.Feature selection — four methods compared, keeping features that scored highly in at least two:
4.statsmodels Logit coefficient significance
5.mlxtend exhaustive 1-feature search
6.SelectKBest (f_classif for continuous, chi2 for categorical)
7.Decision Tree and Random Forest feature importances
8.Modelling — scaled Logistic Regression (deployed) and a Multi-Layer Perceptron (compared). Identical random_state=0 split as the notebook for reproducibility.

Results

Final 8-feature scaled Logistic Regression vs the 13-feature unscaled baseline (test set: 76 patients):

Metric	Baseline (13 features)	Final (8 features, scaled)
Accuracy	0.8158	0.8289
Precision	0.8929	0.8710
Recall	0.6944	0.7500
ROC-AUC	0.8785	0.8861

The final model uses fewer features, scales every feature so convergence is dramatically faster, and improves on the baseline across accuracy, recall, and ROC-AUC — at a small cost in precision.

Selected features

Feature	Why kept
`num_major_vessels`	Strongest discriminator across every selection method; largest scaled coefficient.
`thallium_scint`	High chi² + univariate accuracy; clinically the closest non-invasive proxy for CAD.
`exer_ind_angina`	Strong binary predictor; small confidence interval in univariate test.
`chest_pain_type`	High chi² score; the four-level ordinal captures angina character well.
`st_slope`	Significant chi²; complements ST depression to characterise the stress-test response.
`st_depression`	Significant logistic coefficient; differs sharply between CAD and non-CAD distributions.
`max_heart_rate`	Significant logistic coefficient and high SelectKBest score for continuous variables.
`sex`	Smaller effect but consistently above the noise floor across methods.

Features dropped (age, rest_blood_press, cholesterol, fasting_blood_sugar, rest_ecg) either failed every test or had their signal absorbed by a correlated retained feature (e.g. age ↔ maxheartrate).

Run locally

bash

git clone https://github.com/tejaswirajgit/Coronary-Artery-Disease-Feature-Selection.git
cd Coronary-Artery-Disease-Feature-Selection
pip install -r requirements.txt
python -m src.train         # writes models/model.pkl, models/scaler.pkl, models/metrics.json
streamlit run app.py

Project structure

text

.
├── app.py                       # Streamlit app (HF Spaces entry point)
├── data/
│   └── processed_cleveland.data
├── models/
│   ├── model.pkl                # Trained Logistic Regression
│   ├── scaler.pkl               # StandardScaler fit on train split
│   └── metrics.json             # Test-set metrics + coefficients
├── notebooks/
│   └── coronary-artery-disease-feature-selection.ipynb
├── src/
│   ├── __init__.py
│   ├── preprocess.py            # Cleaning + feature selection
│   └── train.py                 # Reproducible training pipeline
├── .github/workflows/train.yml  # CI: retrains on every push to main
├── requirements.txt
├── LICENSE
└── README.md

Tech stack

—Modelling: scikit-learn (LogisticRegression, StandardScaler, traintestsplit)
—Data: pandas, numpy
—App: Streamlit
—Persistence: joblib
—Plotting (sidebar): matplotlib
—Notebook-only extras: statsmodels, mlxtend, seaborn, Graphviz (for tree visualisations)
—CI: GitHub Actions

Limitations

—Only 303 patients; ROC-AUC of 0.89 on 76 test samples carries non-trivial uncertainty.
—Single-centre, single-era data (Cleveland Clinic, late 1980s) — no guarantee of external validity for other populations or modern clinical protocols.
—The model exposes correlations, not causation.
—Class labels were binarised — severity information (originally 0–4) is discarded.
—Not validated for clinical use under any circumstances.

License

MIT. The Cleveland dataset is provided by the UCI Machine Learning Repository under their own terms.