Stroke Prediction (Supervised Classification)

Overview

This project applies machine learning to predict the likelihood of a stroke based on individual health and demographic factors. The model provides insights that can help identify high-risk individuals potentially aiding early intervention and public health planning.

Objectives

• Gain preliminary insights on the factors influencing stroke in individuals through EDA
• Preprocess and prepare the dataset for accurate model training.
• Develop and compare multiple machine learning models for prediction.
• Evaluate models using reliable performance metrics.
• Understand the main factors contributing to stroke occurrence.

The Dataset

The dataset used is from Kaggle’s Stroke Prediction Dataset and includes a mix of demographic, health, and lifestyle attributes.
The target variable is stroke (1 = stroke occurred, 0 = no stroke)

Feature Categories

• Demographic Feature

  • gender: Male, Female, Other
  • age
  • ever_married: No, Yes
  • work_type: Private, Govt_job, Self_employed, children, Never_worked
  • Residence_type: Urban or Rural

• Health Features

  • hypertension: (0 - no hypertension, 1 - otherwise)
  • heart_disease: (0 - no heart_disease, 1 - otherwise)
  • avg_glucose_level
  • bmi

• Lifestyle features

  • smoking_status: never smoked, formerly smoked, smokes, unknown

  Workflow

1. Exploratory Data Analysis:

   Feature	Finding
   Age	Stroke patients were older on average (≈71 years) than non-stroke patients (≈43 years).
   Avg Glucose Level	Stroke patients had a higher average glucose (105.22 vs. 91.47).
   BMI	The average patient is slightly overweight. Not much difference in stroke rate
   Gender	Males had a slightly higher stroke rate (5.1%) than females (4.7%).
   Hypertension	Stroke rate is over 3x higher among hypertensive patients (13.3%).
   Heart Disease	Stroke rate is 4x higher for those with heart disease (17.0%).
   Ever Married	Married individuals show higher stroke rates (6.6%).
   Work Type	Stroke rate was highest among self-employed individuals (7.9%) and nearly non-existent among those who never worked or were children.
   Residence Type	Urban residents have slightly higher stroke rates (5.2%).
   Smoking Status	Former smokers are most prone (7.9%), possibly due to long-term health impact.

2. Preprocessing

   • Handled missing values with KNNImputer
   • Scaled numeric features
   • Encoded categorical features
   • Addressed target class imbalance with over-sampling technique (SMOTE)

3. Modelling & Evaluation: Implemented and compared the following models based on F1 Score

   • K-Nearest Neighbors
   • Logistics Regression
   • Decision Trees
   • Random Forest
   • Gradient Boosting Classifier Best Model: Logistic Regression: Good generalisation without overfitting. Precision - 16.9% | Highest Recall - 72.5% | Best F1 Score - 27.5%

4. Threshold Tuning: Due to the class imbalance, the standard probability threshold of 0.5 was adjusted to t = 0.30 to improve detection of stroke cases.

   Metric	After Threshold = 0.30
   Precision	⬇️ 13.3%
   Recall	⬆️ 88.7%
   F1 Score	⬇️ 23.2%

5. Model Interpretation Since Logistic Regression does not provide feature importance like tree-based models, interpretability was achieved using Feature Coefficient Analysis which helps to understand the direction and magnitude of each feature.
   The model highlights age , glucose level, hypertension, heart disease, and smoking as key risk factors. In contrast, active employment and non-smoking status act as protective factors.

6. Future Work

   • Implement SHAP for in-depth feature explanations
   • Train deep learning models for further improvements
   • Deploy Streamlit for interactive stroke risk prediction.