HeartStrokePrediction

Introduction

This project aims to analyze and predict the likelihood of a stroke based on various health factors using Machine Learning (ML) and Deep Learning (DL) models.
We performed detailed data cleaning, visualization, feature engineering, model building, and evaluation.
This was done as part of an academic project to understand end-to-end Machine Learning pipelines and Deep Learning modeling.

---

1. Data Preprocessing

Loading and Exploring Data

• The dataset healthcare-dataset-stroke-data.csv was loaded using pandas.
• Basic information such as shape and statistical description was obtained to understand the dataset.

data.shape
data.describe()

---

Handling Missing Values

• The bmi column had missing values.
• Missing bmi values were filled with the mean value of BMI, grouped by gender, marital status, and age group to ensure contextual relevance.
• The id column was dropped as it does not provide any predictive value.

---

Feature Engineering

• An age_group column was created to categorize individuals as Infant, Child, Adolescent, Young Adult, Adult, or Old Aged based on their age.
• Rows with ambiguous gender ("Other") were removed to maintain binary classification.

---

2. Data Visualization and Analysis

We used Plotly Express, Seaborn, and Matplotlib to visualize various feature relationships with stroke:

• Gender vs Stroke (Bar plot)

• Age vs Stroke (Scatter plot)

• Hypertension vs Stroke (Bar plot)

• Heart Disease vs Stroke (Bar plot)

• Marital Status vs Stroke (Bar plot)

• Work Type vs Stroke (Pie chart)

• Residence Type vs Stroke (Bar plot)

• Average Glucose Level vs Stroke (Scatter plot)

• BMI vs Stroke (Scatter plot)

• Smoking Status vs Stroke (Pie chart)

These visualizations helped understand which features were more influential.

---

3. Handling Class Imbalance

The dataset was highly imbalanced (very few people had strokes).

We used upsampling and SMOTE (Synthetic Minority Over-sampling Technique) to balance the dataset:

• Upsampling involved duplicating minority class samples.
• SMOTE generated synthetic samples to balance the classes.

---

4. Data Scaling

• Important continuous features like age, avg_glucose_level, and bmi were normalized using MinMaxScaler or StandardScaler for consistent scaling across ML models.

---

5. Machine Learning Models

We implemented and trained several ML models:

Model	Key Details
Extra Trees Classifier	Ensemble method
Random Forest Classifier	Ensemble method, tuned with hyperparameters
XGBoost Classifier	Advanced boosting technique
Gradient Boosting Classifier	Boosted ensemble method

Each model was:

• Trained on the training set.
• Evaluated on both training and testing sets.
• Assessed using Accuracy Scores, Classification Reports, and Confusion Matrices.

Example model training:

etc_model = ExtraTreesClassifier()
rfc_model = RandomForestClassifier(n_estimators=29, max_leaf_nodes=900, max_features=0.8, criterion='entropy')
xgb_model = XGBClassifier(objective="binary:logistic", eval_metric="auc")
gbc_model = GradientBoostingClassifier(max_depth=29, min_samples_leaf=4, min_samples_split=13, subsample=0.8)

models = [etc_model, rfc_model, xgb_model, gbc_model]

for model in models:
    model.fit(x_train, y_train)

Confusion matrices were plotted for each model.

---

6. Deep Learning Models (ANN)

To enhance prediction accuracy, Deep Learning models were built using TensorFlow and Keras.

Basic Neural Network

• A simple ANN model with two hidden layers.
• Compiled with adam optimizer and binary_crossentropy loss.
• Trained for 50 epochs.

model = Sequential([
    Dense(16, input_dim=X.shape[1], activation='relu'),
    Dense(8, activation='relu'),
    Dense(1, activation='sigmoid')
])

---

Improved Neural Network

• Handled class imbalance using SMOTE.
• Added Dropout layers to prevent overfitting.
• Used EarlyStopping callback to terminate training early if no improvement was observed.

model = Sequential([
    Dense(32, activation='relu', input_dim=X_train.shape[1]),
    Dropout(0.3),
    Dense(16, activation='relu'),
    Dense(1, activation='sigmoid')
])

early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

This improved model showed better generalization and achieved higher accuracy.

---

7. Evaluation

• Classification reports showed precision, recall, f1-score for both classes (Stroke / No Stroke).
• Confusion matrices visualized True Positives, True Negatives, False Positives, and False Negatives.
• Test Accuracy was printed at the end for both ML and DL models.

---

8. Conclusion

Through this project, we learned how to:

• Preprocess messy real-world healthcare datasets.
• Handle missing values thoughtfully.
• Balance imbalanced datasets using resampling and SMOTE.
• Build, train, and evaluate multiple machine learning models.
• Build deep learning models using Keras and TensorFlow.
• Evaluate models using statistical and visual metrics.

This end-to-end project helped us understand the critical steps in building a reliable prediction system for sensitive applications like healthcare.

---

📊 Technologies Used

• Python
• Pandas, Numpy
• Matplotlib, Seaborn, Plotly Express
• Scikit-learn
• XGBoost
• TensorFlow, Keras
• imbalanced-learn (SMOTE)

---

📁 Dataset

• Healthcare Dataset for Stroke Prediction (available on Kaggle).

---