Stroke Prediction Project

Overview

Dataset

• Source: Kaggle Stroke Prediction Dataset
• Features:
  • gender, age, hypertension, heart_disease, ever_married, work_type, Residence_type, avg_glucose_level, bmi, smoking_status
• Target Variable: stroke (Binary: 1 = Stroke, 0 = No Stroke)
• Dataset Size:
  • Original: 5110 rows and 12 columns
  • After Preprocessing (Normal): 4909 rows and 11 columns (after dropping the id column and removing missing values)
  • After Preprocessing (Oversampling with SMOTE): 9400 rows (balanced dataset with 4700 Stroke and 4700 No Stroke cases)
  • Training Data: 7520 rows, 11 columns
  • Test Data: 1880 rows, 11 columns

Workflow Pipeline

1. Data Pre-processing and Preparation

• The id column was removed as it does not contribute to predictive analysis.

• The bmi column had 201 missing values (3.9% of the data). These rows were removed due to their small proportion.

• The dataset was heavily imbalanced, with a 19:1 ratio between No Stroke and Stroke.
• SMOTE Oversampling was applied to balance the dataset:
  • Before SMOTE:
    • No Stroke: 4700 samples
    • Stroke: 209 samples
  • After SMOTE:
    • No Stroke: 4700 samples
    • Stroke: 4700 samples

2. Exploratory Data Analysis

2.1 Correlation Analysis with Target Variable

• Analyzed the correlation between features and the target variable to identify key predictors.
  • Strongest Predictor: age (0.23): The higher the age, the more likely a stroke is.
  • Other Notable Predictors:
    • hypertension (0.14): Patients with hypertension have an amplified risk of stroke.
    • avg_glucose_level (0.14): High glucose levels increase the likelihood of a stroke.
    • heart_disease (0.14): Patients with heart disease face a heightened risk of stroke.

• A heatmap of feature correlations revealed rather obvious correlations, such as:
  • age and ever_married (-0.68): Younger individuals are less likely to be married.
  • age and work_type (-0.42): Younger individuals are less likely to be employed.
  • age and smoking_status (-0.39): Smoking status decreases with younger age.

3. Feature Engineering

• Categorical features were transformed into numerical values for processing:
  • gender: Male (0), Female (1), Other (2)
  • ever_married: Yes (0), No (1)
  • work_type: Private (0), Self-employed (1), Govt_job (2), Children (3), Never_worked (4)
  • smoking_status: Formerly smoked (0), Never smoked (1), Smokes (2), Unknown (3)
  • Residence_type: Urban (0), Rural (1)

• Min-Max scaling was applied to normalize predictors to a range of 0-1, ensuring consistent feature ranges.

4. Model Development and Evaluation

• Before training the models, the dataset was split into training and testing sets:
  • Training Set: 7520 samples (80% of the data) used for training the models.
  • Testing Set: 1880 samples (20% of the data) held out for evaluating the models' performance on unseen data (see confusion matrices).

• Five machine learning models suitable for solving classfication tasks were trained and evaluated:
  1. Random Forest
  2. Logistic Regression
  3. K-Nearest Neighbors (KNN)
  4. Support Vector Machine (SVM)
  5. Naive Bayes

• Each model was evaluated using:
  • Accuracy
  • Precision
  • Recall
  • F1-Score
  • ROC-AUC
  • Runtime (s)
  • Confusion Matrices

Results

Training Metrics

• Accuracy: 82.8%, meaning 82.8% of training samples were classified correctly.
• Precision: 82.9%, indicating that most predicted strokes were true strokes.
• Recall: 82.8%, meaning the model identified 82.8% of actual stroke cases correctly.
• F1-Score: 82.8%, showing a balanced trade-off between precision and recall.
• ROC-AUC: 91.0%, meaning a strong ability to differentiate between stroke and no-stroke cases.
• Runtime: 0.208 seconds, indicating fast training.

Testing Metrics (Confusion Matrices):

• True Positive (Stroke): It correctly predicted 94.6% of all Stroke cases as Stroke.
• True Negative (No Stroke): It correctly predicted 88.9% of all No Stroke cases as No Stroke.
• False Positive (No Stroke predicted as Stroke): 11.1% of No Stroke cases were incorrectly predicted as Stroke.
• False Negative (Stroke predicted as No Stroke): 5.4% of Stroke cases were incorrectly predicted as No Stroke.

Confusion Matrices: After vs. Before Oversampling

Model Interpretability

Feature Importance (Random Forest)

Feature Importance Scores (Random Forest)

Observations and Insights

• Achieved perfect training metrics, along with a reasonable runtime of 0.529 seconds.
• The perfect training scores and the slight degradation in test performance indicated overfitting
• However its generalizability was still the best, as it correctly identified 94.6% of Stroke cases and 88.9% of No Stroke cases on unseen test data (see confusion matrix).

• Random Forest identified the following features as most significant for predicting strokes:
  • age (almost 40% importance): Older individuals are at a higher risk of strokes.
  • avg_glucose_level (second-most important): Higher glucose levels indicate a greater risk.
  • bmi (third-most important): High BMI levels also correlate with a higher stroke risk.

• Confusion matrices revealed that most models failed to identify Stroke cases, with predictions heavily biased toward the majority class No Stroke.
• After oversampling, models showed improved performance in detecting Stroke cases almost equally well correctly prediciting Stroke and No Stroke on unseen data.