🧠 Stroke Prediction Using Machine Learning

A project for CSL2050 - Pattern Recognition and Machine Learning at IIT Jodhpur that develops machine learning models to predict stroke occurrences based on patient demographics, lifestyle, and health features. Our analysis across multiple clinical scenarios highlights the importance of context-specific model selection, with different algorithms showing strengths in different healthcare settings.

---

💡 Problem Statement

Stroke is a leading cause of death and disability worldwide. Early detection is critical for timely intervention. However, predicting stroke is challenging due to the severe class imbalance—stroke cases are rare (19.44:1 ratio in our test set). Our goal is to build a system that:

• Handles imbalanced data using techniques like SMOTE
• Uses feature engineering, selection, and classification models
• Evaluates performance across different clinical scenarios
• Recommends optimal models for specific healthcare contexts

---

📁 Dataset

• Source: Healthcare Dataset Stroke Data on Kaggle
• Raw Data: 5,110 patient records
• Target Column: stroke (1 = stroke, 0 = no stroke)

📊 Features

• gender, age, hypertension, heart_disease, ever_married, work_type, Residence_type, avg_glucose_level, bmi, smoking_status
• Engineered Features: age_group, bmi_category, glucose_category, age_bmi_interaction, hypertension_heart_disease

---

⚙️ Data Preprocessing

• Cleaning: Dropped id, imputed bmi with median
• Encoding: LabelEncoded categorical features
• Feature Engineering:
  • age_group: Young, Middle-Aged, Senior, Elderly
  • bmi_category: Underweight, Normal, Overweight, Obese
  • glucose_category: Low, Normal, Prediabetes, Diabetes, High Risk
  • Interaction terms
• Scaling: Standardized numerical features
• Class Imbalance Handling: Applied SMOTE to 70% training data while preserving natural distribution in 30% test set

⚠️ Imbalance

• Stroke: 249 cases (4.87%) in full dataset
• No Stroke: 4,861 cases (95.13%) in full dataset
• Test set imbalance ratio: 19.44:1

---

📊 Exploratory Data Analysis

• Stroke cases more common among:
  • Married individuals (5.97%)
  • Self-employed (7.8%)
  • Former smokers (7.63%)
• Positive correlations:
  • age & stroke: 0.25
  • hypertension & heart_disease: 0.13

Notebooks available:

• 01_exploratory_data_analysis.ipynb
• 02_data_preprocessing.ipynb
• 03_model_comparison.ipynb

---

🧠 Models Implemented and Analyzed

We implemented and optimized seven machine learning algorithms:

1. Logistic Regression:

   • PowerTransformer normalization, L1 regularization, 'liblinear' solver
   • Best parameters: {'classifier__solver': 'liblinear', 'classifier__penalty': 'l1', 'classifier__class_weight': 'balanced', 'classifier__C': 0.1}
   • Best overall performer (3 out of 4 scenarios)

2. Random Forest:

   • PowerTransformer, SelectKBest feature selection
   • Best parameters: {'feature_selection__score_func': f_classif, 'classifier__n_estimators': 100, 'classifier__criterion': 'entropy', 'classifier__max_depth': 10}
   • Consistent second-best performer

3. Artificial Neural Network (ANN):

   • MLPClassifier with SMOTE balancing
   • Best parameters: {'feature_selection__k': 'all', 'ann__hidden_layer_sizes': (256,), 'ann__activation': 'relu', 'ann__alpha': 0.01, 'ann__learning_rate': 'constant'}
   • Strong performance in limited resource settings

4. Support Vector Machine (SVM):

   • RBF kernel, feature selection (k=25)
   • Best parameters: {'feature_selection__k': 25, 'classifier__kernel': 'rbf', 'classifier__gamma': 0.1, 'classifier__C': 100}
   • Good generalization but lower recall

5. K-Nearest Neighbors (KNN):

   • RobustScaler, mutual information feature selection
   • Best parameters: {'knn__weights': 'distance', 'knn__n_neighbors': 11, 'knn__metric': 'minkowski'}
   • Moderate performance across scenarios

6. Decision Tree:

   • Mutual information feature selection
   • Best parameters: {'feature_selection__percentile': 85, 'classifier__max_depth': 5, 'classifier__criterion': 'log_loss'}
   • Second-best for high-risk screening

7. Gaussian Naïve Bayes (GNB):

   • StandardScaler, mutual information feature selection
   • Best parameters: {'preprocessor__num__scaler': StandardScaler(), 'gnb__var_smoothing': 1e-11}
   • Highest sensitivity (0.80), ideal for high-risk screening

---

📊 Model Performance Across Clinical Scenarios

General Hospital Screening

1. Logistic Regression (Score: 0.7086)
2. Random Forest (Score: 0.6888)
3. Bayesian (Score: 0.6158)
4. Decision Tree (Score: 0.6119)
5. ANN (Score: 0.6040)
6. KNN (Score: 0.5881)
7. SVM (Score: 0.5668)

High-Risk Patient Screening

1. Bayesian (Score: 0.8128)
2. Decision Tree (Score: 0.8092)
3. Logistic Regression (Score: 0.7479)
4. Random Forest (Score: 0.7100)
5. KNN (Score: 0.6296)
6. ANN (Score: 0.5436)
7. SVM (Score: 0.4805)

Limited Resource Setting

1. Logistic Regression (Score: 0.5941)
2. Random Forest (Score: 0.5916)
3. ANN (Score: 0.5852)
4. SVM (Score: 0.5722)
5. KNN (Score: 0.4960)
6. Bayesian (Score: 0.4198)
7. Decision Tree (Score: 0.4154)

Balanced Clinical Decision Support

1. Logistic Regression (Score: 0.5532)
2. Random Forest (Score: 0.5339)
3. ANN (Score: 0.4770)
4. Bayesian (Score: 0.4404)
5. SVM (Score: 0.4335)
6. KNN (Score: 0.4197)
7. Decision Tree (Score: 0.4102)

---

🧪 Key Performance Metrics

Due to the class imbalance, we focused on balanced metrics:

Logistic Regression:

• Accuracy: 0.8584
• Sensitivity: 0.6000
• ROC AUC: 0.8332
• Precision: 0.1940
• Specificity: 0.8717

Bayesian Model:

• Sensitivity: 0.8000
• NPV: 0.9833
• Highest recall for high-risk scenarios

All model metrics are saved in:

• model_metrics.json
• Prediction history in prediction_history.json

---

📌 Recommendations

Based on our analysis, we recommend:

1. For Emergency Screening (High Sensitivity Required): → Use Bayesian Model (0.80 sensitivity, 0.9833 NPV)

2. For General Hospital Screening: → Use Logistic Regression (balanced sensitivity/specificity)

3. For Limited Resource Settings: → Use Logistic Regression (efficient, good specificity)

4. For Balanced Clinical Decision Support: → Use Random Forest (good ROC AUC, interpretable)

---

🚀 Running the Web App

💻 Local Setup

git clone <repo-url>
cd stroke-prediction-project/demo
pip install -r requirements.txt
python app.py

Then go to http://localhost:5000

🧩 App Features

• Model selection based on clinical scenario
• Home page for input
• Displays top risk factors
• Visual results and prediction history
• Robust error handling and logging (app.log)

Templates used:

• index.html, result.html, about.html, error.html

---

🛠️ Project Structure

stroke-prediction-project
├── data/
│   ├── raw/healthcare-dataset-stroke-data.csv
│   └── processed/
│       ├── cleaned_dataset.csv
│       ├── X_test_processed.csv
│       ├── X_train_resampled.csv
│       ├── y_test.csv
│       ├── y_trained_resampled.csv
│       └── feature_statistics.txt
├── demo/
|   └── data/
│       ├── prediction_history.json
|   └── model/
│       ├── model_metrics.json
│       └── stroke_model.pkl
|   └── templates/
│       ├── about.html
│       ├── error.html
│       ├── index.html
│       └── result.html
│   └── app.py
│   └── requirements.txt
├── notebooks/
|   ├── 01_exploratory_data_analysis.ipynb
|   ├── 02_data_preprocessing.ipynb
│   └── 02_model_comparison.ipynb
├── reports/
│   │   ├── mid_reports
|   |   ├── final_report
│   │   └── images....
├── src/
│   ├── models/
│   │   ├── ann.py
|   |   ├── ann_model.pkl
│   │   ├── bayesian.py
|   │   ├── bayesian_model.pkl
│   │   ├── decision_tree.py
|   │   ├── decision_tree_model.pkl
│   │   ├── knn.py
|   │   ├── knn_model.pkl
│   │   ├── logistic_regression.py
|   │   ├── logistic_regression_model.pkl
│   │   ├── random_forest.py
|   │   ├── rf_model.pkl
|   │   ├── svm_model.pkl
│   │   └── svm.py
│   └── data_processing/
│       └── cleaning.py
└── README.md

---

📌 Future Improvements

• Develop ensemble systems combining logistic regression and Bayesian models
• Implement adaptive threshold selection based on clinical context
• Explore deep learning models with attention mechanisms
• Conduct external validation across diverse patient populations
• Integrate with electronic health record systems
• Perform prospective clinical trials to assess real-world impact

---

👥 Contributors

• Anuranjani (B23EE1083)
• Pratheeksha Bangarapu (B23EE1058)
• Polimetla Eshikha (B23CS1053)
• Sai Pragathi Lagudu (B23CM1021)
• Pragna Sree Muvva (B23CM1025)

---

📚 References

• Scikit-learn
• Imbalanced-learn
• Kaggle Dataset
• MLPClassifier - sklearn docs
• PowerTransformer - sklearn docs
• For understanding Pipeline With Hyperparameter Tuning
• GridSearch
• ColumnTransformer