Heading for motor imagery brain-computer interfaces (MI-BCIs) usable out-of-the-lab: Impact of dry electrode setup on classification accuracy

A primary challenge to make motor-imagery Brain-Computer Interfaces (MI-BCIs) technologies usable and actually used out-of-the-lab consists of providing EEG systems that are efficient in terms of classification accuracy- and easy to install, e.g., using a minimal number of dry electrodes. We hypothesize that the optimal signal processing method might depend on the number of (dry) electrodes that are used. Therefore, we compared for the first time the classification accuracy associated with different dry electrode setups, i.e., 7 configurations from 8 to 32 channels, and various signal processing methods, namely (1) regularized Common Spatial Pattern (rCSP) + Linear Discriminant Analysis, (2) rCSP + Support Vector Machine (SVM), (3) Minimum Distance to Riemannian Mean and (4) SVM in the Riemannian Tangent Space. This offline comparison was performed on the data of 10 participants (one session each). Our results suggest that for all methods. MI-BCI performance drops significantly for 8 and 12 channels ( $p ). Moreover, method 3 was associated with the lowest performances ( $p ). Finally, post-doc analyses suggest that methods 1 and 2 perform best with the highest numbers of electrodes 28 and 32. For method 4 the best performance is obtained using 20 and 24 channels, which seems to be the optimal combination ( $p ). These results show the importance of selecting the signal processing pipeline as a function of the location and number of the electrodes.