Handling Class Imbalance in Electroencephalography Data Using Synthetic Minority Oversampling Technique

Abstract

The issue of class imbalance is very common in medical diagnosis applications. The problem of class imbalance arises when the distribution of samples across the known classes is skewed. It affects the performance of robust machine learning algorithms. The performance evaluation parameters used to measure the machine learning models are also very important in the case of class imbalanced datasets. This paper critically evaluates the performance of random forest classifier for the application of seizure state recognition using electroencephalography (EEG) data. Synthetic Minority Oversampling Technique (SMOTE) is used to balance the EEG dataset. Parameters such as accuracy, sensitivity, specificity, precision, false positive rate (FPR), f1-score, and area under the receiver operating characteristic curve (AUC) are considered for evaluation and analysis of the performance of actual class imbalanced dataset and class balanced dataset using SMOTE. The results indicate that the effect of balancing class does not affect the accuracy of the model much, however, it improves the sensitivity and AUC parameters significantly; which are the important parameters in case of measuring the predictive performance of class imbalanced dataset.