Filter Bank Convolutional Neural Network for Short Time-Window Steady-State Visual Evoked Potential Classification.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are increasingly used in Brain-Computer Interfaces (BCIs) for Steady-State Visual Evoked Potential (SSVEP) detection.
• Traditional methods often rely on frequency-domain features (e.g., FFT) or time-domain signals, but frequency-domain analysis can struggle with short time windows and phase information.

Purpose of the Study:

• To propose and evaluate a time-domain-based CNN (tCNN) method for SSVEP detection.
• To further enhance performance in short time windows using a Filter Bank tCNN (FB-tCNN).
• To compare FB-tCNN against Canonical Correlation Analysis (CCA) and other CNN methods.

Main Methods:

• Developed a time-domain CNN (tCNN) that utilizes raw time-domain signals as input.
• Introduced a Filter Bank tCNN (FB-tCNN) to improve feature extraction within short time windows.
• Evaluated FB-tCNN on both a custom dataset and a public SSVEP dataset, comparing it with CCA and other CNN approaches.

Main Results:

• FB-tCNN demonstrated superior performance in short time-window SSVEP detection compared to CCA and other CNN methods.
• Achieved high accuracy rates at a 0.2s time window: 88.36% on the custom dataset and 77.78%/79.21% on the public dataset.
• Investigated the impact of training data size and length, showing FB-tCNN's potential for inter-individual BCI implementation.

Conclusions:

• FB-tCNN offers a significant improvement for SSVEP detection, particularly in scenarios requiring rapid responses.
• Deep learning methods like FB-tCNN present a more straightforward implementation for asynchronous BCI systems compared to CCA.
• The proposed FB-tCNN method shows promise for developing more efficient and robust brain-computer interfaces.