Stroke Patients Analysis And Classification

The project includes analysis of stroke patients dataset and creating classification models for categorizing patients into stroke patients and no-stroke patients.

The project is divided into three stages EDA, Visualization and Modeling ,while in last part I share my thoughts. The data for the project was taken from Kaagle.

1. EDA

1. In this stage, data was analysed to see if there are any null values, data inconsistencies etc. and get rid of it.
2. Only 'BMI' column contained null values which were replaced with mean BMI to maintain the distribution.
3. Column 'gender' had only one value Other, which was replaced with Male value based on values of other features.
4. Also, distribution of data and some basic relationships were visualized to understand the data better.

2. Visualization

1. The new data created from the visualization was used to create a responsive dashboard to compare or analyse patients with respect to lifestyle health features provided in data.
2. Some numerical variables were converted in categorical groups for better categorising.
3. Some new measures were created using DAX to help with visualization and understanding of patients stats better.

3. Modeling

1. The data was preprocessed in the stage with handling categorical variables, Scaling the data etc.
2. The target variable 'stroke' contained heavy imbalance between the categories. The imbalanced was removed using SMOTE.
3. 6 classification models using Decision Tree, KNN, Logistic Regression, Random Forest, SVC, XGBoost were created.
4. The performance was analysed for individual classifier and the best model was selected using hyperparameter tuning.
5. The best model of each classifier was trained and tested on the data and metrics scores were recorded and stored for final comparison and conclusion.
6. The model should be selected based on domain knowledge as the evaluation criteria might change.

4. Conclusion

1. The original data was heavily imbalanced, which affected the result significantly. Using SMOTE improved results slightly which os still good.
2. Regarding Evaluation criteria, In healthcare domain, recall is as important as accuracy, because we want more patients to get treated while avoiding unnecessary treatment for healthy people.
3. The situation demands balanced evaluation between accuracy and recall, which makes 'f1-score with recall score' most valuable metric.
4. Based on above metric, I recommend Logistic Regression. Even though it lacks a little in terms Accuracy. The recall score, f1 score and ROC-AUC score makes it suitable candidate for me. But, it just my opinion.
5. As I mentioned earlier, domain experts can have different opinion depending how they want to handle the problem.