How low can you go: evaluating electrode reduction methods for EEG-based speech imagery BCIs

Speech imagery brain-computer interfaces (SI-BCIs) aim to decode imagined speech from brain activity and have been successfully established using non-invasive brain measures such as electroencephalography (EEG). However, current EEG-based SI-BCIs predominantly rely on high-resolution systems with 64 or more electrodes, making them cumbersome to set up and impractical for real-world use. In this study, we evaluated several electrode reduction algorithms in combination with various feature extraction and classification methods across three distinct EEG-based speech imagery datasets to identify the optimal number and position of electrodes for SI-BCIs. Our results showed that, across all datasets, the original 64 channels could be reduced by 50% without a significant performance loss in classification accuracy. Furthermore, the relevant areas were not limited to the left hemisphere, widely known to be responsible for speech production and comprehension, but were distributed across the cortex. However, we could not identify a consistent set of optimal electrode positions across datasets, indicating that electrode configurations are highly subject-specific and should be individually tailored. Nonetheless, our findings support the move away from extensive and costly high-resolution systems toward more compact, user-specific setups, facilitating the transition of SI-BCIs from laboratory settings to real-world applications.