Predicting Stroke Diagnosis Using Hyperparameter Tuning

This study aims to develop a stroke risk prediction model using a dataset from Kaggle that includes demographic, clinical and lifestyle factors. The process begins with data collection, followed by data exploration to understand the structure, features and existing challenges. The next stage is data preprocessing and cleaning, including handling missing values, coding categorical variables, and mitigating outliers using log normalization and winsorization. The prediction model uses a Random Forest Classifier with hyperparameter tuning via Grid Search. Class balancing was carried out by oversampling to achieve a balanced distribution between stroke sufferers and non-stroke sufferers. Model evaluation shows that although training accuracy remains constant, testing accuracy decreases, indicating overfitting. However, with the addition of estimators, model accuracy increases and accuracy variance decreases, resulting in a more reliable and consistent model. The results show that handling outliers and balancing the dataset significantly improves model performance. The resulting model is more stable and provides more accurate predictions, although it requires improvements in processing time efficiency. It is hoped that these findings will contribute to the development of better stroke prediction models in the future, as well as increasing understanding of the factors that influence stroke in Indonesia.