How Low Can You Go: Evaluating Electrode Reduction Methods for EEG-based Speech Imagery BCIs

Maurice Rekrut^*Johannes IhlTobias JungbluthAntonio Krüger

• Cognitive Assistants, German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany

Speech Imagery Brain Computer Interfaces (SI-BCIs) aim to decode imagined speech from brain activity and have been successfully established even with non-invasive brain measures such as electroencephalography (EEG). However, current EEG-based SI-BCIs mainly rely on highresolution devices with 64 + electrodes, making them cumbersome to setup and inconvenient for real-world use. In this work, we evaluated a set of electrode reduction algorithms in combination with different feature extraction and classification methods on 3 different EEG-based speech imagery datasets to find the best suitable number and position of electrodes for SI-BCI. We show that for all datasets the original 64 channels could be reduced by 50% without a significant performance loss in classification accuracy and that the relevant areas do not necessarily target only the left hemisphere widely known to be responsible for speech production and comprehension, but are spread over the cortex. However, in terms of a best set of specific electrode positions, we cannot draw a consistent conclusion between datasets, suggesting that the setups are highly subject-specific and need to be tailored to the individual. Still, this work suggests reconsidering currently extensive and expensive high-resolution setups and switching to more compact and user-specific ones, to bring SI-BCIs from the lab to real-world application.