CC BY
36
Section BRAIN-COMPUTER INTERFACES, COGNITIVE NAVIGATION WORKSHOP AND NEUROENGINEERING
EEG when 8 participants played it with their gaze only. Moves in the game were made in "control-on" mode of the game with fixations exceeding 500 ms threshold. In the "control-off" mode, fixations did not led to actions, and 500 ms or longer spontaneous fixations were collected. A special procedure was developed to make sure that the analyzed EEG intervals were not contaminated by the artifacts related to eye movement.
The EEG during controlling but not spontaneous fixations showed pronounced negativity in the posterior cortical areas starting early in the course of fixations. Using a simple feature extraction algorithm, greedy feature selection strategy and a linear classifier committee, we obtained, on average, a better than 35% true positive rate for controlling fixations while keeping the false alarm rate at about 10% on the test data with 5-fold cross-validation, much above the random level. More elaborated feature extraction algorithms are currently being tested.
Moreover, a two-threshold strategy was developed to enable smooth interaction even with the current relatively low true positive rate. When a fixation exceeds the first, short (e.g., 500 ms) threshold, the BCI is applied to detect the intention to act on the fixated location. If the BCI misses the intention, the user still may issue the command by continuing fixating the same position until the second (e.g., 1000 ms) threshold is exceeded. Since spontaneous fixations of this length are rare, it is safe to execute a command at this time even without confirmation from the BCI; alternatively, a confirmation from the BCI can be required again but with a low BCI threshold. With such a strategy, the users may develop a more stable EEG pattern associated with controlling fixations, because this will lead to faster move execution.
Our results imply that the "eye-brain-computer" interfaces (EBCIs) can not only helping neurorehabilitation, as the typical BCIs (Kaplan, 2016), but also can be used by healthy persons. Fast converting of intentions into computer actions without using any supplemental tasks (such as computer mouse manipulation, as well as special mental imagery or attention to external stimulation for activating a BCI) may make certain tasks involving interaction with computers especially fluent. This will open new perspectives for unfolding the full scale of benefits from augmenting brain function with the power of computers.
At the conference, these prospects will be discussed within a more general framework of our current and planned studies, including the search for new intention markers both in the EEG and the magnetoencephalogram (MEG), investigating the factor of the feeling of agency and free will in the use of BCIs, and online interaction of the users with different types of BCIs and EBCIs.
Acknowledgements
Parts of this work related to the specific methods of intention marker detection, their use in the EBCIs and the studies of feeling of agency and free will were supported by the Russian Science Foundation, grant 14-28-00234.
References
1. A. One, B. Two, and C. Three, Phys. Rev., 1972, 8(3), 555-566.
2. D. C. Engelbart, Augmenting Human Intellect: A Conceptual Framework, 1962.
3. B. M. Velichkovsky, and J. P. Hansen, SIGCHI Conf. Human Factors in Comp. Systems, 1996, 496-503.
4. J. Proteak, K. Ihme, and T. O. Zander, UAHCI. Design Methods, Tools, and Interaction Techniques for eInclusion, 2013, 662-671.
5. A. Y. Kaplan, Fiziologiya Cheloveka [Human Physiol.], 2016, 42(1), 118-127 (in Russian).
BCI Matrix Speller Based on Coded Visual Evoked Potentials
R.K. Grigoryan1 *, D.B. Flatov1, A. Ya. Kaplan1,2,3
1 Lomonosov Moscow State University, Moscow, Russian Federation;
2 Lobachevsky State University of Nizhni Novgorod, Nizhni Novgorod, Russian Federation;
3 Pirogov National Russian Medical University, Moscow, Russian Federation. * Presenting e-mail: grraph.bio@gmail.com
Aims of the study
Brain-computer interface (BCI) is a system that utilizes neurophysiological correlates of attention to establish communication with computer. The most popular type of BCIs is BCI based on visual evoked potentials (VEP BCI), for example P300 BCI or steady-state VEP BCI. Here we examine another paradigm - BCI based on code-modulated visual evoked potentials (C-VEP BCI). Within this approach, m-sequence [1] is used as pattern of visual stimulation. The crucial feature of such sequence is its autocorrelation function that has only one peak, and equals zero in all posi-
OM&P
OM&P
Section BRAIN-COMPUTER INTERFACES, COGNITIVE NAVIGATION WORKSHOP AND NEUROENGINEERING
tions except this peak. Circular shift is introduced, and resulting shifted sequences are used to determine flash pattern of LEDs or elements of a computer screen. When user gazes at the elements, flashing this pattern, code-modulated response is generated [2]. This response can be extracted using canonical correlation analysis, and used to identify the target that has been chosen by user. C-VEP BCI combines the best characteristics of T-VEP and F-VEP BCI: it doesn't require long training and it is possible to introduce many different commands at time, thus enhancing ITR.
The aim of this work was to examine differences in the evoked potentials and BCI performance in response to different stimulation sequences.
Methods
20 healthy adults (10 males and 10 females) with normal or corrected-to-normal vision participated in the experiment. Visual stimuli were presented on a LCD monitor with refresh rate 120 Hz. 32 targets were arranged as a matrix. Targets were labeled with Cyrillic letters. Each target was altering between black and white with pattern derived from 64bit binary m-sequence. The period of the sequence was 1 second. Two flashing modes were present: "straight pattern" ("1" bit of m-sequence denotes white state of stimuli, "2" denotes black), and "inverse pattern", derived from the same m-sequence by inverting it (XORed with 64 zeros). EEG was recorded from 8 channels: Pz, P1, P2, PO1, PO2, PO3, PO4, POz at 500 Hz. Learning session was followed by online mode, where participant tried to select all 32 targets. Classification algorithm based on CCA was employed in online mode, classifying EEG samples of variable length, up to 5 seconds long.
ȧ 0.5 iif u
Time, seconds
Fig.1. Correlation coefficient between single trial EEG in response to target number 25 and average EEG response from learning session, visualized over time. Red color depicts maximum correlation coefficient
Conclusions
We have successfully created BCI based on C-VEP paradigm. Its characteristics in terms of number of commands, ITR and accuracy make this type of interfaces a viable replacement for traditional and well-proven t-VEP P300 BCIs. The main weakness of C-VEP spellers is visual environment, which is too aggressive and may lead to operator tiredness. The nature of differences between evoked patterns proposes complex mechanism of Ep generation.
Acknowledgements
This study was partially supported by funding from Russian Science Foundation (#15-19-20053). References
1. Wolfmann, J. IEEE Transactions on Information Theory, 1992, 38(4), 1412-1418.
2. Bin, G., Gao, X., Wang, Y., Li, Y., Hong, B., & Gao, S. , Journal of Neural Engineering, 2011, 8(2)
