Cardiovascular Disease Prediction: A Machine Learning Approach

Table of Contents

• Project Overview
• Dataset
• Pipeline Summary
• Technical Stack
• Model Overview
• Evaluation & Metrics
• Key Findings
• Project Structure
• Installation
• Results
• Model Comparison
• Clinical Implications
• Future Work
• References
• Author

---

Project Overview

Cardiovascular diseases represent the leading cause of death worldwide, accounting for approximately 19.8 million deaths annually according to the World Health Organization (WHO). Early identification of individuals at risk is essential for implementing preventive strategies and improving long-term patient outcomes.

This project develops a complete machine learning pipeline for predicting the presence of cardiovascular disease using clinical patient data. It demonstrates:

• Data preprocessing with missing value imputation, outlier detection (Z-score method), and encoding of categorical variables
• Comprehensive exploratory analysis of clinical features and their relationships
• Feature engineering including MinMax scaling and Random Forest-based feature selection
• Comparative evaluation of four classification algorithms (Logistic Regression, Gaussian Naive Bayes, SVM, Random Forest)
• Performance assessment using metrics suited for medical applications
• Fully reproducible workflow with saved models and preprocessing artifacts

The pipeline provides a clear example of designing and evaluating classifiers for medical diagnostic support, with particular attention to the sensitivity-specificity trade-off critical in healthcare applications.

---

Dataset

Source: UCI Machine Learning Repository (Heart Disease Dataset)

The dataset combines patient records from four independent studies:

• Cleveland Clinic Foundation
• Hungarian Institute of Cardiology
• University Hospital Zurich, Switzerland
• VA Medical Center, Long Beach

Dataset Characteristics:

• Total Samples: 920 patient records (908 after outlier removal)
• Features: 13 clinical and diagnostic attributes
• Target Classes: 0 (absence) to 4 (presence with varying severity levels) → Binary classification (0 = Healthy, 1-4 = Heart Disease)

Feature Description

Feature	Type	Description	Clinical Significance
age	Numerical	Patient age in years	Progressive risk factor
sex	Binary	1 = male, 0 = female	Men have higher baseline risk
cp	Categorical	Chest pain type (1-4)	Symptom indicator
trestbps	Numerical	Resting blood pressure (mm Hg)	Hypertension marker
chol	Numerical	Serum cholesterol (mg/dL)	Lipid profile
fbs	Binary	Fasting blood sugar > 120 mg/dL	Diabetes indicator
restecg	Categorical	Resting ECG results	Cardiac electrical activity
thalach	Numerical	Maximum heart rate achieved	Exercise response
exang	Binary	Exercise-induced angina	Ischemia indicator
oldpeak	Numerical	ST depression during exercise	Ischemia severity
slope	Categorical	ST segment slope	Stress response pattern
ca	Numerical	Number of major vessels colored	Coronary obstruction
thal	Categorical	Thalassemia type	Perfusion status

---

Pipeline Summary

---

Technical Stack

Category	Libraries
Data Handling	pandas, NumPy, SciPy
Modeling	scikit-learn
Feature Selection	Random Forest Feature Importance
Evaluation & Metrics	scikit-learn, Matplotlib, Seaborn
Visualization	matplotlib, seaborn
Reproducibility & Export	joblib, pickle

---

Model Overview

Models Implemented

Model	Description	Optimal Hyperparameters
Logistic Regression	Linear classifier with probability outputs	C=10, penalty='l2', solver='liblinear'
Gaussian Naive Bayes	Probabilistic classifier assuming feature independence	var_smoothing=1e-9
SVM (RBF kernel)	Margin-based classifier with non-linear boundary	C=1, kernel='rbf'
Random Forest	Ensemble of decision trees	max_depth=5, n_estimators=100, min_samples_split=5

Why These Models?

• Logistic Regression: Interpretable baseline, provides probability estimates
• Gaussian Naive Bayes: Fast, works well with continuous features, handles uncertainty
• SVM: Effective in high-dimensional spaces, flexible decision boundaries
• Random Forest: Robust to outliers, captures non-linear relationships, provides feature importance

---

Evaluation & Metrics

Given the medical context, special attention was paid to metrics that matter in clinical settings:

Metric	Formula	Clinical Relevance
Accuracy	(TP + TN) / (TP + TN + FP + FN)	Overall correctness of predictions
Recall (Sensitivity)	TP / (TP + FN)	Critical: Ability to detect diseased patients (minimize false negatives)
Macro F1-Score	Average of class-wise F1 scores	Harmonic mean of precision and recall, averaged equally across both classes
MCC	(TP×TN - FP×FN) / √[(TP+FP)(TP+FN)(TN+FP)(TN+FN)]	Balanced measure robust to class imbalance, ranging from -1 to +1
ROC-AUC	Area Under ROC Curve	Threshold-independent measure of discriminative ability

Key Evaluation Techniques

• Confusion Matrices: Visual assessment of prediction patterns
• ROC Curves: Threshold-independent performance visualization
• Cross-validation: 5-fold stratified CV for hyperparameter tuning with scoring='roc_auc'
• Macro-averaged metrics: Equal weight to both classes

---

Key Findings

Clinical Insights from EDA

Age and Disease Severity: Clear progression — healthy patients cluster at 45–65 years, severe cases predominantly occur in individuals over 65.

Gender Disparity:

• 78% of men have heart disease vs 12% of women
• Men develop disease 7–10 years earlier and progress to severe stages more frequently

Chest Pain: Asymptomatic patients show 79% disease probability, highlighting that absence of symptoms does not exclude disease.

ST Depression: Strongest numerical predictor — healthy patients cluster near zero while disease patients show elevated values.

Feature Importance (Top 10 Selected)

Rank	Feature	Importance	Clinical Significance
1	chol	0.156	Cholesterol drives atherosclerosis
2	thalach	0.125	Exercise capacity reflects cardiac function
3	age	0.115	Cumulative cardiovascular risk
4	oldpeak	0.096	Ischemia indicator
5	exang	0.087	Exercise-induced angina
6	cp_atypical angina	0.082	Atypical symptom presentation
7	trestbps	0.076	Hypertension marker
8	thal	0.056	Perfusion defects
9	sex_male	0.051	Demographic risk factor
10	cp_non-anginal pain	0.043	Non-cardiac chest pain

---

Project Structure

HeartDisease_ML/
│
├── datasets/                                 # Raw dataset files
│   ├── processed.cleveland.data
│   ├── processed.hungarian.data
│   ├── processed.switzerland.data
│   ├── processed.va.data
│   └── heart_disease_combined.csv           # Merged and cleaned dataset
│
├── models/                                   # Trained models and artifacts
│   ├── logistic_regression.joblib
│   ├── naive_bayes.joblib
│   ├── svm.joblib
│   ├── random_forest.joblib
│   ├── scaler.joblib                         # Fitted MinMaxScaler
│   └── feature_names.txt                     # Selected features
│
├── images/                                   # All visualizations from the notebook
│   ├── EDA/
│   │   ├── target_distribution.png
│   │   ├── demographic_analysis.png
│   │   ├── heatmap_cp.png
│   │   ├── heatmap_ecg.png
│   │   ├── heatmap_slope.png
│   │   └── boxplot.png
│   ├── evaluation/
│   │   ├── naive_bayes.png
│   │   ├── logistic_regression.png
│   │   ├── random_forest.png
│   │   └── svm.png
│   ├── figures/
│   │   ├── Coronary-heart-disease.jpg
│   │   └── visual_pipeline.png
│   └── feature_importance.png
│
├── project/
│   ├── HeartDisease_Prediction_Report.pdf
│   └── heart_disease.ipynb                   # Complete analysis pipeline
│
├── results/                                   # Evaluation outputs
│   ├── model_comparison.csv                   # Metrics comparison table
│   └── best_models_per_metric.csv             # Best model for each metric
│
├── .gitignore
└── README.md

---

Installation

Prerequisites

• Python 3.8 or higher
• pip package manager

Setup

1. Clone the repository

git clone https://github.com/sofianatale/HeartDisease_ML.git
cd HeartDisease_ML

2. Create a virtual environment (recommended)

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

pip install pandas numpy matplotlib seaborn scikit-learn scipy joblib

4. Run the Jupyter Notebook

jupyter notebook project/heart_disease.ipynb

---

Results

Final Test Set Performance

Model	Accuracy	Recall	Macro F1	MCC	ROC-AUC
Random Forest	0.8516	0.9300	0.8470	0.7036	0.9094
Logistic Regression	0.8516	0.8900	0.8490	0.7000	0.9110
Gaussian Naive Bayes	0.8516	0.8800	0.8500	0.7000	0.9100
SVM	0.8462	0.9100	0.8420	0.6902	0.9130

Best Model by Metric

Metric	Best Model	Value
Accuracy	Logistic Regression / Random Forest / GNB	0.8516
Recall (Sensitivity)	Random Forest	0.9300
Macro F1	Gaussian Naive Bayes	0.8500
MCC	Random Forest	0.7036
ROC-AUC	SVM	0.9130

---

Model Comparison

Key Observations

• Random Forest demonstrates the best balance for clinical applications, achieving the highest Recall (0.93) and MCC (0.7036), making it the most suitable model for minimizing false negatives
• SVM shows the highest discriminative ability (ROC-AUC = 0.9130) and strong recall (0.91), but at the cost of lower MCC due to more false positives
• Gaussian Naive Bayes achieves the best Macro F1 (0.8500), indicating excellent balance between precision and recall across both classes
• Logistic Regression provides consistent performance across all metrics, making it a reliable baseline
• All models demonstrate strong predictive capability with ROC-AUC scores exceeding 0.90

Sensitivity-Specificity Trade-off

Model	Recall (Sensitivity)	Specificity	Clinical Trade-off
Random Forest	0.93	0.7683	Best for screening (minimizes missed diagnoses)
SVM	0.91	0.7317	High sensitivity but more false alarms
Logistic Regression	0.89	0.8049	Balanced approach
Naive Bayes	0.88	0.8171	Conservative (fewer false positives, more false negatives)

---

Clinical Implications

Model Selection Rationale

For medical screening applications, sensitivity (recall) is often prioritized over specificity, as missing a disease diagnosis has more severe consequences than false alarms. Based on this criterion:

• Random Forest (sensitivity = 0.93, MCC = 0.7036) → Best choice for screening
• SVM (sensitivity = 0.91) → Good alternative with excellent ROC-AUC
• Logistic Regression (sensitivity = 0.89) → Balanced option with good specificity
• Naive Bayes (sensitivity = 0.88) → Conservative approach, minimizes false positives

Why Random Forest is recommended as the final model:

• Highest Recall (0.93) — identifies 93% of actual disease cases
• Highest MCC (0.7036) — best overall correlation between predictions and reality
• Strong ROC-AUC (0.9094) — excellent discriminative ability
• Competitive accuracy — ties for best at 0.8516
• Robust to outliers and non-linear relationships — ensemble nature handles complex medical data

Risk Factor Confirmation

The feature importance analysis confirms established medical knowledge:

• Cholesterol is the dominant predictor
• Exercise capacity (thalach) strongly indicates cardiac function
• Age remains a fundamental risk factor
• ST depression validates as key diagnostic indicator

---

Future Work

• External Validation: Test on independent clinical datasets
• Multi-class Prediction: Extend to severity levels (0–4) for risk stratification
• Deep Learning: Compare with neural networks for potential improvements
• Explainable AI: Implement SHAP/LIME for individual prediction explanations
• Clinical Deployment: Develop web interface or API for healthcare integration
• Feature Engineering: Create composite clinical indicators (e.g., rate-pressure product)
• Longitudinal Analysis: Incorporate time-series data for progressive risk assessment

---

References

Dataset & Problem Domain

1. UCI Heart Disease Dataset
2. Ali MM, Paul BK, Ahmed K, Bui FM, Quinn JMW, Moni MA. Heart disease prediction using supervised machine learning algorithms: Performance analysis and comparison. Comput Biol Med. 2021 Sep;136:104672.

Machine Learning Models

3. Nick TG, Campbell KM. Logistic regression. Methods Mol Biol. 2007;404:273-301.
4. Hu J, Szymczak S. A review on longitudinal data analysis with random forest. Brief Bioinform. 2023 Mar 19;24(2):bbad002.
5. Lai Z, Chen X, Zhang J, Kong H, Wen J. Maximal Margin Support Vector Machine for Feature Representation and Classification. IEEE Trans Cybern. 2023 Oct;53(10):6700-6713.
6. Pavan Venkata, Vivek Pandya. Data mining model and Gaussian Naive Bayes based fault diagnostic analysis of modern power system networks. Materials Today: Proceedings, Volume 62, Part 13, 2022, Pages 7156-7161.

Evaluation Metrics

7. Chicco D, Tötsch N, Jurman G. The Matthews correlation coefficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Min. 2021 Feb 4;14(1):13.
8. Obuchowski NA, Bullen JA. Receiver operating characteristic (ROC) curves: review of methods with applications in diagnostic medicine. Phys Med Biol. 2018 Mar 29;63(7):07TR01.
9. Hicks SA, Strümke I, Thambawita V, Hammou M, Riegler MA, Halvorsen P, Parasa S. On evaluation metrics for medical applications of artificial intelligence. Sci Rep. 2022 Apr 8;12(1):5979.

Clinical Background

10. World Health Organization (2025). Cardiovascular diseases (CVDs)
11. Goldsborough E 3rd, Osuji N, Blaha MJ. Assessment of Cardiovascular Disease Risk: A 2022 Update. Endocrinol Metab Clin North Am. 2022 Sep;51(3):483-509.

---

Author

Sofia Natale

• Affiliation: AML-BASIC 2025 – University of Bologna
• Email: sofia.natale@studio.unibo.it
• GitHub: sofianatale

For questions, collaborations, or feedback, please open an issue or reach out via email.