Stroke-Risk-Prediction-Using-Machine-Learning

Predictive analysis of brain stroke risk factors and model development using Python, SQL, and ML techniques.

Table of Contents

1. Overview
2. Dataset
3. Installation
4. Usage
5. Methodology
   • Data Cleaning & Preparation
   • Exploratory Data Analysis (EDA)
   • Statistical Testing
   • Feature Engineering
   • Modeling
6. Results
7. Contributing
8. License

---

Overview

This project analyzes a publicly available stroke dataset to identify key risk factors and build predictive models for 30-day stroke occurrence. We apply data cleaning, statistical tests, visualization, and machine learning pipelines (Logistic Regression, Decision Tree, Random Forest, XGBoost) to evaluate model performance and interpretability.

Dataset

• Source: Kaggle: Brain Stroke Dataset
• Records: 4,981
• Features:
  • Demographics: gender, age, ever_married, residence_type, work_type, smoking_status
  • Health: hypertension, heart_disease, avg_glucose_level, bmi
  • Target: stroke (0 = no stroke, 1 = stroke)

Installation

1. Clone the repository:

   git clone https://github.com/Hemanth-072/Stroke-Risk-Prediction-ML.git
   cd Stroke-Risk-Prediction-ML
   
    Create a virtual environment and install dependencies:
   
    python3 -m venv venv
    source venv/bin/activate        # Linux/macOS
    # .\venv\Scripts\activate      # Windows
    pip install -r requirements.txt
   
    Place brain_stroke.csv in the project root (or modify the path in the notebooks/scripts).
   

Usage

Jupyter Notebooks:

    01_Data_Cleaning_EDA.ipynb – Data loading, cleaning, EDA, outlier handling

    02_Statistical_Analysis.ipynb – Normality tests, chi-square, Mann–Whitney U, Kruskal–Wallis

    03_Modeling.ipynb – Preprocessing pipelines, model training, evaluation, and comparison

Scripts (optional):

python src/train.py         # Runs end-to-end preprocessing and model training
python src/predict.py       # Serves a saved model for inference

Methodology Data Cleaning & Preparation

Standardize column names and data types

Confirm and handle missing values (none in this dataset)

Winsorize outliers (1st–99th percentile) for continuous features

Exploratory Data Analysis (EDA)

Histograms & box plots for age, avg_glucose_level, bmi

Count plots for categorical variables by stroke outcome

Correlation matrix for numeric and binary features

Statistical Testing

Normality: Shapiro–Wilk on capped continuous features

Non-parametric: Mann–Whitney U (continuous vs. stroke)

Categorical: Chi-square tests for independence

Multi-group: Kruskal–Wallis for distributions across stroke groups

Feature Engineering

One-hot encoding for categorical variables

Train-test split (80/20 stratified)

SMOTE to address class imbalance

Modeling

Baseline Logistic Regression (with class_weight='balanced')

Decision Tree, Random Forest, and XGBoost classifiers

Hyperparameter tuning via RandomizedSearchCV

Evaluation metrics: ROC-AUC, PR-AUC, accuracy, precision, recall, F1-score, confusion matrices

Model interpretability: SHAP / feature‐importance plots

Results

Key Risk Factors: Age, hypertension, heart disease, marital status, work type, and smoking status all showed statistically significant associations with stroke risk (p < 0.05).

Best Model: Random Forest and XGBoost achieved ROC-AUC ≈ 0.85 on the held-out test set.

Interpretability: Feature importance analyses highlighted age and hypertension as top predictors.

For full metric tables, plots, and model outputs see the notebooks in this repo.