Individual-finger motor imagery classification: a data-driven approach with Shapley-informed augmentation.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Classifying individual five-finger motor imagery (MI) using electroencephalography (EEG) is challenging due to signal variability.
• Existing methods struggle with inconsistent temporal features in EEG data across sessions for the same subject.

Purpose of the Study:

• To systematically assess the classification of individual five-finger MI using single-trial time-domain EEG signals.
• To identify neural correlates discriminating between five-finger MI tasks.
• To introduce and evaluate a Shapley-informed augmentation technique for improving within-subject classification accuracy.

Main Methods:

• Data-driven analysis of time-domain EEG signals for within-subject and cross-subject MI classification.
• Model explainability analysis to identify neural correlates.
• Implementation and evaluation of Shapley-informed augmentation to address temporal feature inconsistencies.

Main Results:

• Neural information discriminating individual five-finger MI was found in the parietal cortex.
• Shapley-informed augmentation significantly improved classification accuracy (average 26.3%±6.70) in sessions with inconsistent temporal features.
• Average classification accuracies achieved were approximately 60.0% (within-session), 50.0% (within-subject), and 40.0% (leave-one-subject-out).

Conclusions:

• The parietal cortex contains discriminative neural information for individual five-finger MI.
• Shapley-informed augmentation effectively mitigates temporal variability, enhancing BCI application accessibility.
• This research advances personalized BCI systems by addressing EEG signal inconsistencies in motor imagery classification.