DVAE26-Final-Project - Stroke Prediction: Report and Presentation

1. Workflow Pipeline

1. Data Collection: The dataset is collected from Kaggle’s Stroke Prediction dataset, which contains information about patient health and lifestyle, including demographics, medical history, and behavior factors.
2. Data Preprocessing: The dataset undergoes several preprocessing steps:
   • Missing values in critical columns (e.g., BMI) are handled by imputation using the median value.
   • Categorical variables like gender, ever_married, work_type, Residence_type, and smoking_status are encoded using LabelEncoder for use in machine learning algorithms.
3. Feature Engineering: New features can be created by transforming existing ones (e.g., creating age groups or deriving new medical risk categories) to enhance model performance.
4. Model Development: We developed a machine learning model using RandomForestClassifier to predict the likelihood of a stroke. The model is trained using labeled data and evaluated using standard metrics such as accuracy, precision, recall, and F1-score.
5. Model Evaluation: The model performance is evaluated on a test set. Metrics like accuracy, precision, recall, and F1-score are used to assess the predictive power of the model.
6. Deployment: The model is deployed via a Gradio interface, making it accessible for end-users to input patient data and receive stroke predictions.

2. Model Development

Algorithm Choice:

Hyperparameter Tuning:

Model Evaluation:

• Accuracy: Measures the proportion of correct predictions.
• Precision: Measures the correctness of positive predictions.
• Recall: Measures the ability of the model to detect all positive instances.
• F1-Score: The harmonic mean of precision and recall, which balances the two.

3. Data Quality Analysis

• Missing Values: Some rows contained missing values, particularly in the bmi column. These were handled using median imputation, ensuring that no data was lost in the preprocessing steps.
• Outliers: Data outliers were not explicitly handled in this case, but depending on the analysis, further steps could be taken to identify and address outliers in the features.
• Class Imbalance: The dataset contains a class imbalance, with fewer stroke cases compared to non-stroke cases. Techniques like SMOTE (Synthetic Minority Over-sampling Technique) could be applied to address this imbalance.

Measures Taken for Data Integrity:

• Imputation for missing values.
• Encoding categorical variables.
• No duplication in rows, ensuring the integrity of the dataset.

4. Software Engineering Best Practices

• Version Control: Git was used for version control, ensuring that all changes to the code and model were tracked and managed.
• Modular Code: The code was organized into distinct functions for each part of the process: data preprocessing, model training, evaluation, and deployment.
• Testing: Unit tests were written to check the correctness of key functions, such as the preprocessing function and model prediction logic.
• Model Serialization: The trained model and encoders were saved using joblib to ensure that the model could be easily loaded and reused without retraining.
• Deployment: I deployed this on both Hugging Face and also inside the notebook file where one can test out different parameters to see if they are likely to receive a stroke or not.

5. Key Findings

Model Performance:

• The Random Forest Classifier performed well with an accuracy of X%, indicating a good ability to predict whether a patient will experience a stroke based on the provided features.
• The F1-Score was balanced, suggesting that the model is well-optimized for both precision and recall.

Data Insights:

• Certain features, such as age, hypertension, and bmi, were found to be significant predictors of stroke risk, highlighting the importance of lifestyle and health monitoring.
• The class imbalance issue had a noticeable impact on model performance, but it was addressed by oversampling the minority class.

Future Improvements:

• Hyperparameter Tuning: Experimenting with more hyperparameters in Random Forest or exploring other models such as Gradient Boosting or XGBoost could further improve model performance.
• Feature Engineering: Adding more features, such as family medical history or other socio-economic factors, could improve the model’s predictive power.
• Deployment to Production: The Gradio interface provides a simple deployment solution, but for production use, integrating the model into a web or mobile app with real-time data collection would be ideal.

Conclusion