Electroencephalography (EEG) plays a crucial role in clinical neurodiagnostics, particularly in epileptic focus localization and deep sleep detection. However, the limited availability of annotated EEG data hinders the generalization capability of deep learning models. This study proposes a unified EEG classification framework that applies three lightweight data augmentation techniques, namely time shifting, amplitude scaling, and noise addition, to enrich training diversity and enhance model robustness. The framework is evaluated using DeepConvNet, ShallowConvNet, and EEGNet on two public datasets that represent physiological and pathological EEG tasks. Experimental results show that data augmentation consistently improves classification performance across all models and tasks. Importantly, even when baseline accuracies are already high, the proposed augmentation strategies provide additional gains of up to approximately 2.06% in deep sleep detection and 4.07% in epileptic focus localization. These findings demonstrate that simple augmentation methods can effectively improve the robustness and classification performance of EEG-based deep learning models, especially under data-limited conditions.
Chen et al. (Sun,) studied this question.