Reference Bank Multi-Feature Extraction for EEG-Based Concentration Discrimination

Jin Woo Choi, Jaehoon Choi, Sungho Jo

Abstract: Discriminating concentration of a user is one of the few tasks that non-invasive BCIs can be applied in real-life situations. To have EEG-based BCIs more accessible to users, attempts have been made in terms of both hardware, where EEG acquisition devices have been redesigned to be more affordable and comfortable to wear, and software, where better algorithms have been introduced to improve the interface’s performance. For concentration discrimination, a task highly relevant to EEG signals from the frontal lobe, using only electrodes in the forehead has previously been proposed to further simplify the setup required for EEG measurement. However, this requires careful selection of ground and reference electrodes; having ground and reference electrodes located on the forehead close to other electrodes results in less discriminant signals, while placing them on mastoids or other EEG neutral locations makes the interface bulky to wear and more susceptible to various sources of artifacts, such as ocular and facial muscle movements. Thus, in this paper, we propose a reference bank multi-feature extraction approach that aims to improve previous existing deep learning based models with multiple forms of re-referenced data. We conducted an experiment using dry electrodes placed only on the forehead to collect brain signals related to concentration and resting state to evaluate our approach. Our method was applied to three pre-existing CNN-based models, exhibiting an average increase of 3.19% in their classification performance.