Selective Subject Pooling Strategy to Improve Model Generalization for a Motor Imagery BCI

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) enable communication for individuals with severe motor impairments.
• BCI systems typically require extensive calibration for personalized classifier training.
• High inter-subject variability in electroencephalogram (EEG) data poses a significant challenge for subject-independent BCI development.

Purpose of the Study:

• To investigate selective subject pooling strategies for enhancing subject-independent motor imagery (MI) BCI performance.
• To compare the effectiveness of pooling all available subjects versus strategically selecting subjects for training data.
• To identify criteria for optimal subject selection in developing robust subject-independent BCIs.

Main Methods:

• Comparative performance testing of different subject pooling strategies using public MI BCI datasets.
• Evaluation of subject-independent BCI classifier accuracy based on selected training data.
• Analysis of inter-subject variability impact on classifier performance.

Main Results:

• Selective subject pooling strategies demonstrated improved performance compared to using all available subjects.
• The proposed strategy effectively mitigated performance degradation caused by inter-subject EEG variability.
• The effectiveness of selective pooling was validated on public MI BCI datasets.

Conclusions:

• Strategic selection of subjects for training data can significantly enhance subject-independent BCI performance.
• This approach offers a viable solution to overcome EEG variability challenges in BCI development.
• Criteria for subject selection are proposed to guide the creation of more effective subject-independent BCIs.