Heart Disease Prediction using Machine Learning (UCI Heart Dataset)

Heart disease remains one of the leading causes of death worldwide, and early risk assessment can guide patients toward timely medical attention. Clinicians use many risk factors — age, blood pressure, cholesterol, ECG readings, exercise response — but weighing them together consistently is difficult by hand. Machine learning is a natural fit: a classifier trained on real patient records can learn how these clinical attributes combine into a risk signal. This project does exactly that on the well-studied UCI Heart Disease (Cleveland) dataset: 303 records, 13 clinical features, one binary target. Students often run this dataset through a library call without understanding the pipeline, so this build emphasizes the full engineering workflow — data cleaning, feature encoding, train/test splitting, model comparison, and evaluation with confusion matrices and ROC curves. The web demo turns the trained approach into something tangible: enter the 13 features and see a risk probability, the risk band, and which factors contributed most.

1. The UCI Heart Disease (Cleveland) dataset — 303 patient records, 13 clinical attributes, binary target — is loaded from the UCI Machine Learning Repository.
2. Missing values are handled and categorical features (chest-pain type, resting ECG, slope, thal) are encoded; numeric features are standardized for the linear model.
3. The data is split 80/20 into train and test sets (stratified, so both classes are represented), with cross-validation on the training portion for tuning.
4. Logistic Regression (L2), Random Forest and XGBoost are trained on the same split; hyperparameters are tuned with grid search on the training folds.
5. Each model is evaluated once on the held-out test set: accuracy, precision, recall, F1-score, ROC-AUC, confusion matrix and per-class report.
6. The best model is serialized with its preprocessing pipeline and wired into the web demo, which computes the risk probability and surfaces the top contributing features for any entered patient profile.

Tech stack:

• Python 3, scikit-learn
• XGBoost
• pandas, NumPy
• Matplotlib, Seaborn (evaluation plots)
• Jupyter Notebook (training & evaluation)
• Flask + HTML/CSS/JavaScript (risk-prediction web demo)
• UCI Heart Disease (Cleveland) dataset

Dataset & model details

• Dataset: UCI Heart Disease (Cleveland) — 303 patient records, 13 clinical attributes (age, sex, chest-pain type, resting BP, cholesterol, fasting blood sugar, resting ECG, max heart rate, exercise angina, ST depression, ST slope, major vessels, thalassemia) plus a binary target (heart disease present/absent: 138 present, 165 absent), from the UCI Machine Learning Repository.
• Task: Binary classification; input = 13 clinical features, output = probability of heart-disease presence.
• Model: Three classifiers compared — Logistic Regression (L2 regularization, interpretable baseline), Random Forest (tree ensemble, feature importances), XGBoost (gradient-boosted trees, typically strongest on this tabular data).
• Metrics: Accuracy, precision, recall, F1-score and ROC-AUC on a held-out test set, plus confusion matrices and ROC curves. Design target ~85–90% test accuracy; the report documents the actually achieved figures after the training run — no accuracy is claimed as measured until then.

• [Risk-prediction web demo] Enter the 13 clinical features (age, sex, chest-pain type, resting BP, cholesterol, fasting sugar, resting ECG, max heart rate, exercise angina, ST depression, slope, vessels, thal) and get an instant risk probability with a gauge, risk band, and the top contributing factors explained.
• [Three-model comparison] Logistic Regression, Random Forest and XGBoost trained side by side in the same notebook, so the report can show which model fits this tabular data best and why.
• [Full training & evaluation notebook] Data loading, missing-value handling, categorical encoding, train/test split, scaling, training, and evaluation in one reproducible Jupyter notebook.
• [Confusion matrix & classification report] Per-class precision, recall and F1, so the viva can discuss exactly where each model errs (false negatives matter most in a medical context).
• [ROC curves & AUC comparison] ROC curves for all three models plotted together with AUC scores — the standard ranking-quality metric for binary medical classifiers.
• [Feature importance analysis] Which clinical attributes drive the predictions, visualized and interpreted for the report.
• [Exported trained models] Saved model files plus the preprocessing pipeline, so the demo scores new patients without retraining.
• [Medical-disclaimer framing] Every output of the demo and report states clearly that this is an academic prototype, not a diagnostic tool — an honest and expected viva discussion point.

• Complete training & evaluation Jupyter notebook (cleaning, encoding, training, tuning, evaluation)
• Trained model files with the preprocessing pipeline
• Risk-prediction web demo (13-feature form, probability gauge, risk band, factor explanations)
• Confusion matrices, ROC curves and feature-importance plots
• Project report PDF (background, dataset description, methodology, model comparison, results, medical limitations)
• PPT presentation for final review
• Viva Q&A preparation document (classification metrics, overfitting, why ensembles, medical-AI limits)