Stroke Prediction Analysis 🚨

A comprehensive pipeline for predicting stroke risk using real-world clinical data, with an emphasis on handling severe class imbalance and optimizing clinically relevant metrics.

---

📋 Table of Contents

1. Project Overview
2. Dataset
3. Pipeline & Workflow
4. Insights
5. Usage Examples
6. Contributing
7. License
8. Contact

---

📌 Project Overview

Predicting stroke risk in patients is critical for early intervention. This project:

• Utilizes a real-world Stroke Prediction dataset from Kaggle.
• Tackles severe class imbalance (~4% stroke cases).
• Compares multiple classifiers (e.g., Logistic Regression, XGBoost).

---

📊 Dataset

Source: Kaggle Stroke Prediction Dataset

Feature Category	Columns
Demographics	gender, age
Health Metrics	avg_glucose_level, bmi
Medical History	hypertension, heart_disease
Lifestyle & Social	ever_married, work_type, residence_type, smoking_status
Target	stroke (0 = No, 1 = Yes)

After cleaning (dropping missing BMI, removing outliers), 4,909 records remain.

---

🔄 Pipeline & Workflow

1. Data Loading & Cleaning

   • Standardize column names
   • Handle missing BMI values

2. Feature Engineering

   • Encoding for linear models
   • Label Encoding for tree-based models

3. Train-Test Split

   • Stratified (80% train / 20% test)

4. Scaling

   • StandardScaler for distance-based models

5. Imbalance Handling

   • SMOTE in an imblearn.Pipeline (avoids data leakage)

6. Model Training & Tuning

   • KNN, Logistic Regression, Decision Tree, Random Forest, XGBoost
   • GridSearchCV optimizing F1 on stroke class

7. Threshold Tuning

   • Adjust probability cut-offs for high-recall vs balanced use cases

---

Hyperparameter & Threshold Tuning

• GridSearchCV for optimal hyperparameters & SMOTE sampling ratio
• Custom thresholds (e.g., 0.20 vs. 0.50) to target recall or precision based on clinical needs

---

📈 Insights

• Logistic Regression offers the best balance (highest F1 & ROC AUC).
• XGBoost at 0.20 excels in recall—ideal for screening scenarios.

---

💻 Usage Examples

# Load trained model
import joblib
model = joblib.load('gridsearch/xgb_model.pkl')

# Predict on new data
import pandas as pd
new_data = pd.read_csv('csv_files/sample_patient.csv')
pred_probs = model.predict_proba(new_data)[:, 1]
pred_labels = (pred_probs >= 0.2).astype(int)

---

🤝 Contributing

1. Fork the repo
2. Create a feature branch (git checkout -b feature/YourFeature)
3. Commit your changes (git commit -m "Add new analysis")
4. Push (git push origin feature/YourFeature)
5. Open a Pull Request

---

📄 License

This project is licensed under the MIT License. See LICENSE for details.

---

📬 Contact

Authors

Mic-dev-gif (Michele Montalvo)

• Email: michelemontalvo@outlook.com
• GitHub: github.com/Mic-dev-gif

Jorge M. M. L. Rodrigues

• Email: jorgemmlrodrigues@gmail.com
• GitHub: github.com/JorgeMMLRodrigues

Feel free to open issues or discussions!