Improving Classification using IS-SMOTE on Imbalanced Heart Failure Data

Heart failure is a fatal cause of mortality globally. One in every four individuals is afflicted by and succumbs to heart disease. Given the considerable complexity of heart disease classification, this research primarily centers on the classification of heart failure patients and strives to predict heart failure survival with high accuracy. Because of class imbalance problem, the classification result is prone to bias towards the dominant majority class. An existing popular technique for handling class imbalance issues is the Synthetic Minority Oversampling Technique (SMOTE). SMOTE gives equal importance to every minority instance of oversampling. Furthermore, a significant but unresolved problem is the generation of minority synthetic instances, the number of which is fixed at five for each instance. This creates overlapping samples between closer instances. In this research, a modified variant of SMOTE is proposed, namely, Instance Selection based SMOTE (IS-SMOTE), to mitigate the problem of SMOTE. In our proposed IS-SMOTE, oversam- pling rates are not fixed at five, and resembling instances are oversampled based on their distance from other instances. The closest instances are given maximum importance over the furthest ones and generate a smaller number of synthetic instances. The most difficult issues, particularly oversampling rates and figuring out how many neighbors there are for each sample in SMOTE, can be resolved by the proposed IS-SMOTE. ISSMOTE outperforms existing SMOTE in terms of average performance metrics, including Accuracy, Precision, Recall, and F-Score, across five base classifiers: SVM, GNB, RF, DT, and XGBoost, as demonstrated in experiments conducted on heart failure data.