Comprehensive Analysis of Data Augmentation Methods in Classification for an Imbalanced Epilepsy Dataset

Abstract

Imbalanced class distribution reduces the generalizability of classifiers in EEG-based epilepsy detection. This study examines the impact of the synthetic minority oversampling technique (SMOTE) and its variants on imbalanced electroencephalography (EEG) data, utilizing an end-to-end data processing pipeline. Band-limited filtering is applied as pre-processing, and then the training data is gradually oversampled by 20% increments in four scenes. Experiments are conducted on coarse-k-nearest neighbor (Coarse-KNN), bagged trees, and artificial neural network (ANN) classifiers, and evaluation is performed using accuracy, precision, recall, F1 score, and Matthew’s correlation coefficient (MCC) metrics. In Scene #4, where the inter-class imbalance is eliminated, Borderline-SMOTE yielded the highest and most consistent results (F1 Score = 0.903–0.937, MCC = 0.830–0.894). Safe level-SMOTE (SL-SMOTE) and SMOTE/Geometric-SMOTE(G-SMOTE) produced second-ranked results. The findings demonstrate that appropriate variant selection provides consistent gains even across classifiers, making Borderline-SMOTE the recommended approach for imbalanced EEG classification. Furthermore, in the detailed analysis of ensemble sampling limits, SMOTE-based combined approaches (e.g., SL + G SMOTE) also produced consistent results. Basic descriptive statistics (mode, median, variance, and kurtosis) of the synthetic samples were found to be comparable to those of the real data, providing additional evidence of distributional consistency.