Covid-19 detection using dominant SMOTE in imbalance classification

Abstract

Global healthcare systems have faced difficulties since the start of the COVID-19 epidemic. For overburdened hospitals, identifying positive patients is a simple and effective fix. The disproportionate distribution of classes poses a significant challenge in identifying the positive case of COVID-19, leading to biased prediction outcomes favoring dominant classes. Consequently, classifiers struggle to learn from imbalanced datasets, resulting in reduced performance. Various techniques, such as oversampling, undersampling, and hybrid sampling, have been proposed to mitigate this issue. However, the Synthetic Minority Oversampling Technique (SMOTE) remains a commonly utilized resampling method despite its limitations, including class mixture. To address these shortcomings, I introduce Dominant SMOTE, a modified version of SMOTE. The proposed method comprises of developing a dominant sample selection approach based on numerical attribute values from the minority class, and selecting the nearest neighbors from the majority class for each minority class sample based on dominance values to achieve balanced dataset. The proposed method is compared with traditional SMOTE and Out-Layer SMOTE, evaluating accuracy, precision, recall, and F1-score on two benchmark datasets. The results indicate that the proposed model outperforms than both the traditional SMOTE and Out-Layer SMOTE.