Learning Invariant Patterns Based on a Convolutional Neural Network and Big Electroencephalography Data for...

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Traditional brain-computer interfaces (BCIs) require subject-specific calibration, limiting their immediate usability.
• Subject-independent BCIs aim to eliminate this calibration step but face challenges due to electroencephalography (EEG) noise and inter-subject variability.

Purpose of the Study:

• To develop and validate a subject-independent P300 BCI using an invariant pattern learning method.
• To leverage convolutional neural networks (CNNs) and big EEG data for robust feature extraction across subjects.

Main Methods:

• A CNN was trained on a large dataset (200 subjects) of EEG data from a P300 spelling task.
• The model was designed to extract subject-independent features for predicting brain activity.
• Data was collected using two different amplifier types to ensure cross-device generalizability.

Main Results:

• The proposed method demonstrated significant cross-subject and cross-amplifier effects.
• Average accuracy exceeded 80% across nearly all subjects.
• Over 50% of subjects achieved accuracies greater than 85%.

Conclusions:

• The invariant pattern learning method is effective for creating subject-independent P300 BCIs.
• This approach enables high-accuracy BCI operation for a majority of users without requiring individual calibration.
• The findings pave the way for more accessible and user-friendly BCI applications.