Continual-Learning-Enhanced CNN-Transformer Framework for Real-Time Motor-Imagery BCI in Virtual Environments

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI)-based brain-computer interfaces (BCIs) offer intuitive neural interaction but face challenges like long calibration, user variability, and non-stationary EEG signals.
• Dry-electrode EEG enhances convenience but yields noisier signals, complicating real-time, multi-class MI decoding.
• Existing methods struggle with continuous signal drift, limiting long-term BCI performance.

Purpose of the Study:

• To develop a robust real-time four-class MI-BCI framework utilizing dry electrodes.
• To address the challenges of signal non-stationarity and calibration requirements through continual learning.
• To enhance the practical deployment and sustainability of MI-BCIs in intelligent neurotechnology.

Main Methods:

• A hybrid CNN-Transformer model was employed for MI decoding.
• Immersive action observation (AO) in virtual reality was used for pre-training the model.
• An online continual learning strategy adapted the model to evolving dry-EEG features during user interaction.

Main Results:

• The proposed framework demonstrated improved decoding accuracy for four motor classes in real-time.
• The system showed strengthened sensorimotor activation over time, indicating effective adaptation.
• The continual learning approach successfully mitigated performance degradation in extended MI-BCI sessions.

Conclusions:

• The integration of AO and continual learning with a CNN-Transformer model enhances MI-BCI robustness, especially with dry electrodes.
• The framework effectively manages signal drift and reduces calibration burden, enabling user-specific and session-to-session adaptation.
• This approach supports sustainable long-term deployment of MI-BCIs for rehabilitation and intelligent neurotechnology.