A Deep Neural Network for SSVEP-Based Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) enable communication and control through neural signals.
• Target identification in BCI spellers involves classifying electroencephalogram (EEG) signals to predict intended characters.
• Steady-state visually evoked potentials (SSVEP) are commonly used in BCI spellers, where visual stimuli tagged with distinct frequencies elicit characteristic EEG responses.

Purpose of the Study:

• To propose a novel deep neural network (DNN) architecture for enhanced target identification in SSVEP-based BCI spellers.
• To improve the accuracy and information transfer rates (ITRs) of BCI spelling systems.

Main Methods:

• A novel DNN architecture was developed to process multi-channel SSVEP data, utilizing convolutions across sub-bands, channels, and time.
• The DNN was trained in two stages: first, a global model was learned across all subjects, followed by subject-specific fine-tuning.
• The model was evaluated on two large-scale public datasets (benchmark and BETA) comprising 105 subjects and 40 characters.

Main Results:

• The proposed DNN achieved high information transfer rates (ITRs) of 265.23 bits/min and 196.59 bits/min on the benchmark and BETA datasets, respectively.
• These results were obtained with a short stimulation time of only 0.4 seconds.
• The DNN significantly outperformed existing state-of-the-art techniques on both datasets.

Conclusions:

• The developed DNN architecture represents a significant advancement in SSVEP-based BCI spellers, achieving unprecedented accuracy and ITRs.
• The approach demonstrates high performance and applicability to general SSVEP systems.
• This technique holds substantial potential for applications in communication, rehabilitation, and control within BCI systems.