This study investigates the impact of various data-augmentation techniques on the performance of neural networks in EEG-based motor imagery three-class event classification. EEG data were obtained from a publicly available open-source database, and a subset of 25 patients was selected for analysis. The classification task focused on detecting two types of motor events: imagined movements of the left hand and imagined movements of the right hand. EEGNet, a convolutional neural network architecture optimized for EEG signal processing, was employed for classification. A comprehensive set of augmentation techniques was evaluated, including five time-domain transformations, three frequency-domain transformations, two spatial-domain transformations and two generative approaches. Each method was tested individually, as well as in selected two- and three-method cascade combinations. The augmentation strategies were tested using three data-splitting methodologies and applying four ratios of original-to-generated data: 1:0.25, 1:0.5, 1:0.75 and 1:1. Our results demonstrate that the augmentation strategies we used significantly influence classification accuracy, particularly when used in combination. These findings underscore the importance of selecting appropriate augmentation techniques to enhance generalization in EEG-based brain–computer interface applications.
Sztyler et al. (Sat,) studied this question.