BGTransform: a neurophysiologically informed EEG data augmentation framework

Objective. Deep learning has emerged as a powerful approach for decoding electroencephalography (EEG)-based brain-computer interface (BCI) signals. However, its effectiveness is often limited by the scarcity and variability of available training data. Existing data augmentation methods often introduce signal distortions or lack physiological validity. This study proposes a novel augmentation strategy designed to improve generalization while preserving the underlying neurophysiological structure of EEG signals.Approach. We propose Background EEG Transform (BGTransform), a principled data augmentation framework that leverages the neurophysiological dissociation between task-related activity and ongoing background EEG. In contrast to existing methods, BGTransform generates new trials by selectively perturbing the background EEG component while preserving the task-related signal, thus enabling controlled variability without compromising class-discriminative features. We applied BGTransform to three publicly available EEG-BCI datasets spanning steady-state visual evoked potential and P300 paradigms. The effectiveness of BGTransform is evaluated using several widely adopted neural decoding models under three training regimes: (1) without augmentation (baseline model), (2) with conventional augmentation methods, and (3) with BGTransform.Main results. Across all datasets and model architectures, BGTransform consistently outperformed both baseline models and conventional augmentation techniques. Compared to models trained without BGTransform, it achieved average classification accuracy improvements of 2.45%-15.52%, 4.36%-17.15% and 7.55%-10.47% across the three datasets, respectively. In addition, BGTransform demonstrated greater robustness across subjects and tasks, maintaining stable performance under varying recording conditions.Significance. BGTransform provides a principled and effective approach to augmenting EEG data, informed by neurophysiological insight. By preserving task-related components and introducing controlled variability, the method addresses the challenge of data sparsity in EEG-BCI training. These findings support the utility of BGTransform for improving the accuracy, robustness, and generalizability of deep learning models in neural engineering applications.