Improving Classification Performance on Imbalanced Stroke Datasets Using Oversampling Techniques

Stroke is the second leading cause of death worldwide and a major factor in long-term disability. Although early detection based on machine learning continues to be developed, it still faces challenges in the form of data imbalance, which can reduce classification performance. This study aimed to evaluate the effectiveness of several oversampling techniques, such as SMOTE, Borderline-SMOTE, and SVM-SMOTE, in improving the performance of stroke classification models on imbalanced data. The methods used included the application of three oversampling techniques, namely SMOTE, Borderline-SMOTE, and SVM-SMOTE, to balance the data distribution. Furthermore, Random Forest and XGBoost algorithms were used as classification models to identify stroke cases. The results of this study show that the use of oversampling techniques significantly improves model performance, especially in MCC and AUC metrics, compared to models without oversampling. Borderline-SMOTE provides the best results, with the highest accuracy of 96.45% on Random Forest and 96.41% on XGBoost, as well as MCC and AUC values that are consistently superior to other techniques. Furthermore, this study highlights that the use of Borderline-SMOTE significantly enhances model performance, as evidenced by an increase in MCC of 87.51% and an AUC of 45.40% in Random Forest, along with an increase in MCC of 76.52% and an AUC of 41.81% in XGBoost. These findings confirm that Borderline-SMOTE is an effective approach for dealing with data imbalance and improving the model's ability to detect minority classes in stroke classification.