Towards the Correction of Covariate Shift in EEG-Based Passive Brain-Computer Interfaces for Out-of-Lab Applications

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Wearable electroencephalography (EEG) is crucial for passive Brain-Computer Interface (pBCI) systems in Industry 5.0.
• Covariate shift, caused by EEG headset repositioning, significantly degrades classification performance.
• Developing robust pBCI models for real-world applications requires addressing signal variability.

Purpose of the Study:

• To investigate a linear transformation approach for mitigating covariate shift in EEG signals.
• To evaluate the impact of reference and channel changes on classification accuracy.
• To assess the effectiveness of normalizing shifted data using a template.

Main Methods:

• Simulations were performed using various covariate shift conditions.
• A linear transformation function was applied to normalize EEG data.
• Classification performance was evaluated before and after applying the transformation.

Main Results:

• Normalizing covariate shift-affected data with shift-free data as a template significantly improved classification performance.
• Accuracy loss decreased from 14% to 6% in the worst case and 5% to 4% in the best case.
• Improvements were more substantial with larger shifts, involving changes in both reference and channels.

Conclusions:

• The linear transformation method effectively mitigates covariate shift in wearable EEG data.
• This approach enhances the reliability of pBCI systems for out-of-the-lab applications.
• Findings support the development of more robust pBCI models for Industry 5.0 and beyond.