Toward zero-calibration MEG brain-computer interfaces based on event-related fields.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Magnetoencephalography (MEG) provides excellent spatiotemporal resolution for brain activity.
• Practical application of MEG in brain-computer interfaces (BCIs) is hindered by user-specific calibration and inter-subject variability.

Purpose of the Study:

• To develop and evaluate a zero-calibration MEG-based BCI system.
• To leverage spatial filtering and deep learning for robust BCI performance across subjects.

Main Methods:

• An online event-related field (ERF)-based MEG BCI was developed using a visual oddball paradigm.
• xDAWN spatial filtering and a DeepConvNet model were employed for classifying neural responses.
• Leave-one-subject-out (LOSO) cross-validation was used to assess zero-calibration, cross-subject generalization.

Main Results:

• The online BCI achieved 94.29% accuracy and 20.47 bits/min information transfer rate (ITR).
• The zero-calibration approach using xDAWN and DeepConvNet demonstrated 80.32% average accuracy and 12.75 bits/min ITR.
• Successful cross-subject generalization was achieved, indicating plug-and-play capability.

Conclusions:

• Zero-calibration MEG BCIs are feasible for practical applications.
• The combination of spatial filtering and deep learning enhances cross-subject generalization in MEG BCIs.
• This approach reduces the need for individual user calibration, increasing BCI accessibility.