A procedure to minimize EEG variability for BCI applications

A Brain-Computer Interface (BCI) decodes brain activities to translate them into computer commands. Electro-encephalography is the most widely adopted technique for brain signal recording in BCIs, because of practical and safety reasons. However, EEG signals show a significant intra-subject variability, which constitutes a major challenge for BCI development. The main goal of this work is to improve a pseudo-online movement detection system using motor imagery EEG signals that simulate the BCI input. We propose a strategy that aims at minimizing the effects of the poor spatial resolution and the active reference electrode based on finding the best combinations of electrode pairs. The proposed method finds subject-specific pairs of electrodes along with signal transformations that provide the more stable results. The average accuracy across 15 subjects was 95 %. It was also seen that energy signals in the delta band (0-4 Hz) of the electrode line CCP (according to the 10-20 system) are associated to the lowest variability. The hypothesis of lower variability being associated to movement related information and therefore to higher accuracy in classification was confirmed by the results. The main conclusion is that it is possible to overcome in some level the signal variability without introducing mathematical or physical uncertainties inherent to commonly adopted approaches such as spatial filters or volume conduction modeling, for instance. The contribution of this work is the procedure to minimize EEG variability for BCI applications. The significance is the possibility to apply the procedure to any set of channels and transformations.

Automated summary

A Brain-Computer Interface (BCI) translates brain activities into computer commands through decoding brain signals, with electro-encephalography being the most widely adopted technique for signal recording. However, the high intra-subject variability of EEG signals poses a challenge for BCI development. This study aims to improve a pseudo-online movement detection system by using motor imagery EEG signals, proposing a strategy to minimize the effects of poor spatial resolution and active reference electrode by finding the best combinations of electrode pairs. The average accuracy across 15 subjects was 95%.