Short-length SSVEP data extension by a novel generative adversarial networks based framework.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potentials (SSVEPs) are crucial for high-performance brain-computer interfaces (BCIs).
• Current SSVEP-based BCIs face limitations due to extensive user calibration data requirements and data length constraints.
• Generative adversarial networks (GANs) show promise in synthesizing electroencephalography (EEG) data to overcome these limitations.

Purpose of the Study:

• To propose TEGAN, a GAN-based network for extending the time-window length of SSVEP signals.
• To enhance the performance of frequency identification methods in SSVEP-based BCIs, particularly with limited calibration data.
• To reduce the calibration time and cost associated with real-world BCI applications.

Main Methods:

• Developed TEGAN, a GAN-based end-to-end signal transformation network for extending SSVEP signal length.
• Implemented a two-stage training strategy and LeCam-divergence regularization for GAN training.
• Evaluated TEGAN on two public SSVEP datasets (4-class and 12-class).

Main Results:

• TEGAN significantly improved the performance of traditional and deep learning-based frequency recognition methods with limited calibration data.
• The classification performance gap between different frequency recognition methods was narrowed.
• Demonstrated the feasibility of extending short-time SSVEP signals for high-performance BCI development.

Conclusions:

• TEGAN effectively transforms short SSVEP signals into longer artificial ones, enhancing BCI performance.
• The proposed GAN-based approach has significant potential for reducing BCI calibration time and costs.
• TEGAN facilitates the development of more practical and accessible real-world BCI systems.