Stroke Prediction Project

Project Overview

This project focuses on predicting the risk of stroke using patient data. The objective is to develop an effective machine learning model that assists healthcare professionals in identifying high-risk individuals based on various clinical and lifestyle factors.

Technologies Used

• Python (Pandas, NumPy, Scikit-Learn, XGBoost, TPOT)
• Power BI (Dashboard Visualization)
• Jupyter Notebook
• Streamlit (Prototype Web Application)

Dataset Description

• Source: Publicly available health dataset.
• Key Features:
  • Demographics: Age, Gender, Residence Type
  • Health Factors: Hypertension, Heart Disease, BMI, Glucose Levels
  • Lifestyle: Smoking Status, Work Type, Marital Status
• Target Variable: Stroke (0 = No, 1 = Yes)

Exploratory Data Analysis (EDA)

• The dataset is highly imbalanced (only ~5% stroke cases).
• Age, hypertension, heart disease, and glucose levels are significant predictors.
• No strong multicollinearity found among features.
• Visualizations include:
  • Correlation Heatmap
  • Feature Distribution Plots
  • Class Imbalance Visualization

Data Preprocessing

• Removed irrelevant id column.
• Imputed missing BMI values using the median.
• Handled missing values in smoking_status by introducing a new category Unknown.
• Applied One-Hot Encoding for categorical variables.
• StandardScaler used for normalizing numerical features.

Models Applied

1. Logistic Regression

   • Accuracy: 95%
   • Stroke Recall: 2% (Poor recall due to class imbalance)
   • ROC-AUC: 0.84

2. Random Forest Classifier

   • Accuracy: 95%
   • Stroke Recall: 0%
   • ROC-AUC: 0.78
   • Highly biased toward the majority class.

3. XGBoost Classifier

   • Accuracy: ~84%
   • Stroke Recall: 18% (Improved)
   • ROC-AUC: ~0.78

4. TPOT Optimized Pipeline

   • Accuracy: 84%
   • Stroke Recall: 50% (Best Recall Achieved)
   • F1-Score: 0.24
   • This pipeline achieved the best balance between detecting stroke and minimizing false positives.

Final Model Recommendation

• TPOT Optimized Pipeline was selected as the final model due to its superior recall performance, critical in medical risk prediction where missing a positive case is highly undesirable.

Power BI Dashboard

• An interactive dashboard was created to visualize:
  • Class Distribution
  • Feature Importance
  • Risk Profiles based on key features like Age and Glucose Level

Future Enhancements

• Implement SMOTE or other resampling techniques for further balancing the dataset.
• Try advanced ensemble models like LightGBM or CatBoost.
• Improve feature engineering with interaction terms and polynomial features.
• Deploy the model via Streamlit for real-time risk prediction.