Data Augmentation for Subject-Independent SSVEP-BCIs via Simultaneous Spatial-Energy Representation.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Subject-independent classification in electroencephalography (EEG) brain-computer interfaces (BCIs) is crucial for widespread adoption.
• Current deep learning (DL) data augmentation methods struggle to address inter-subject variability in EEG signal styles.

Purpose of the Study:

• To propose a novel data augmentation method, Simultaneous Spatial-Energy Representation (SSER), to improve subject-independent classification in EEG-BCIs.
• To enhance the robustness of DL models against individual-specific style characteristics in EEG signals.

Main Methods:

• SSER utilizes singular value decomposition (SVD) to extract spatial and energy representations from EEG signals.
• These representations are mixed across domains during signal reconstruction to generate diverse styles.
• This approach aims to learn domain-invariant features and improve robustness to style variability.

Main Results:

• SSER outperformed existing data augmentation techniques on public EEG datasets.
• The method demonstrated strong generalization across different DL models.
• Offline and online experiments with 30 subjects confirmed SSER's effectiveness.

Conclusions:

• SSER provides a richer characterization of EEG signal style variability through simultaneous manipulation of spatial and energy representations.
• The method significantly advances subject-independent classification for EEG-BCIs.
• This innovation facilitates broader real-world applications of EEG-based BCIs.