A transformer-based deep neural network model for SSVEP classification.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP) is a key control signal in brain-computer interface (BCI) systems.
• Conventional SSVEP classification methods require subject-specific calibration data, limiting practical applications.
• Developing inter-subject methods for SSVEP classification is crucial for reducing calibration demands.

Purpose of the Study:

• To propose a novel deep learning model, SSVEPformer, for SSVEP classification in an inter-subject scenario.
• To leverage Transformer architecture for enhanced SSVEP signal processing.
• To explore the use of complex spectrum features and filter bank technology for improved classification performance.

Main Methods:

• Proposed SSVEPformer, the first Transformer-based deep learning model for SSVEP classification.
• Utilized complex spectrum features of SSVEP data as model input to capture spectral and spatial information.
• Developed an extended version, FB-SSVEPformer, incorporating filter bank technology to exploit harmonic information.

Main Results:

• The proposed SSVEPformer and FB-SSVEPformer models demonstrated superior classification accuracy and information transfer rates compared to baseline methods.
• Experiments were conducted on two open datasets with varying numbers of subjects and targets.
• The models achieved significant improvements in inter-subject SSVEP classification performance.

Conclusions:

• Deep learning models based on Transformer architecture are feasible and effective for SSVEP data classification.
• The proposed SSVEPformer and FB-SSVEPformer can significantly alleviate the calibration procedure in practical SSVEP-based BCI systems.
• These models represent a promising advancement for more accessible and efficient BCI applications.