Heart Disease & Stroke Diagnosis with Random Forest

Course: Intelligent Systems (CO5134) — Ho Chi Minh University of Technology (VNU-HCM) Instructor: Prof. Quan Thanh Tho Students: Vuong Minh Toan (2491057), Tran Nguyen Huan (2370692) Date: November 2024 Deliverable: Web application (React + Flask) & ML pipeline for diagnostic support

---

📌 Objectives

• Develop a Random Forest model to predict heart disease and stroke based on clinical and lifestyle data.
• Provide a web application for the community: input health indicators → receive predictions + basic recommendations → download reports.

---

👥 Stakeholders & Benefits

• General users: self-assess risk, improve awareness, low cost.
• Community organizations/NGOs: support health campaigns, identify high-risk groups.
• Government/Public health sector: monitor health trends (anonymized), design early interventions.

---

🧩 System Requirements

Functional: input personal data, display predictions + insights, export reports. Non-functional: intuitive UI, response time < 5 seconds, target uptime 99.9%.

---

🏗️ Architecture & Technology

• Frontend: React.js — input forms, results page, responsive design, API integration via Axios.
• Backend: Flask (Python) — load .pkl models, data processing, prediction, return results; modular architecture.
• Overall flow: user input → frontend calls REST API → backend preprocessing & inference → return results → user downloads report.
• Deployment (suggested): React (Netlify/Vercel/S3), Flask (Gunicorn/uWSGI on EC2/Heroku/Azure), monitoring via Sentry/Prometheus/Grafana, analytics with Google Analytics.

---

🗃️ Data

• heart.csv (~1025 samples) — features: age, sex, trestbps, chol, fbs, restecg, thalach, exang, oldpeak, slope, ca, thal, target. target: 1 = heart disease, 0 = no disease.

• stroke.csv (~5110 samples) — features: gender, age, hypertension, heart_disease, ever_married, work_type, residence_type, avg_glucose_level, bmi, smoking_status, stroke. stroke: 1 = stroke, 0 = no stroke.

Key Preprocessing Steps

• Encoding: binary label mapping; One-Hot Encoding for categorical variables (work_type, residence_type, smoking_status…).
• Missing data: ~201 NULL values in bmi (stroke.csv) → dropped due to sufficient dataset size and numeric nature.
• Imbalanced data: stroke dataset is highly imbalanced → applied SMOTE for minority class oversampling.
• Scaling & splitting: train/test split; normalization using StandardScaler and shuffle when needed.

---

🤖 Model & Pipeline

1. Exploration: used LazyPredict to benchmark multiple supervised models → selected Random Forest due to strong performance on both tasks.

2. Training: RandomForestClassifier(random_state=42) + hyperparameter tuning using GridSearchCV (parameters: n_estimators, max_depth, min_samples_split, min_samples_leaf, max_features).

3. Evaluation: Accuracy, Confusion Matrix, Classification Report, ROC/AUC.

Results (from report)

• Heart disease: Accuracy ≈ 0.99, AUC = 1.00 (ROC reaches top-left corner).
• Stroke: Accuracy ≈ 0.98, AUC = 1.00 (very high discriminative performance).

Note: Some reported values in the confusion matrix section are not fully consistent. It is recommended to rerun the notebook/script to extract final metrics from best_estimator_.

---

🔌 API (Draft)

• POST /api/predict/heart → receive health indicators → return heart disease prediction.
• POST /api/predict/stroke → receive health indicators → return stroke prediction.
• GET /api/healthz → check server status.
• POST /api/feedback → collect user feedback (optional).

Backend handles data type conversion, loads .pkl models, performs inference, and returns JSON responses; frontend displays results and allows report download.

---

🧪 How to Run (Suggested)

Backend (Flask)

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# place models at ./models/heart_rf.pkl and ./models/stroke_rf.pkl
export FLASK_APP=app.py
flask run --host 0.0.0.0 --port 8000

Frontend (React)

cd web
npm install
npm run dev  # or npm run build && npm run preview

---

🔒 Ethical & Legal Considerations

• This is a decision-support tool, not a substitute for medical diagnosis. Users are strongly advised to consult healthcare professionals.
• Ensure data anonymization and security compliance with local regulations.

---

🔗 Source Code & Demo

• Source code: above.
• Demo: DEMO.mp4

---

📣 Future Improvements

• Tune decision thresholds to improve recall (reduce false negatives).
• Add class weighting, probability calibration (Platt/Isotonic), and explainability (SHAP).
• Implement data drift monitoring and post-deployment model tracking.

---