Freeze the BCI until the user is ready: a pilot study of a BCI inhibitor
FFreeze the BCI until the user is ready:a pilot study of a BCI inhibitor
L. George, L. Bonnet, A. L´ecuyer
INRIA, Campus Universitaire de Beaulieu, F-35042 Rennes Cedex, France { laurent.f.george, laurent.bonnet, anatole.lecuyer } @inria.fr Abstract
In this paper we introduce the concept of Brain-Computer Interface (BCI) inhibitor, which ismeant to standby the BCI until the user is ready, in order to improve the overall performanceand usability of the system. BCI inhibitor can be defined as a system that monitors user’sstate and inhibits BCI interaction until specific requirements (e.g. brain activity pattern,user attention level) are met. In this pilot study, a hybrid BCI is designed and composed ofa classic synchronous BCI system based on motor imagery and a BCI inhibitor. The BCIinhibitor initiates the control period of the BCI when requirements in terms of brain activityare reached (i.e. stability in the beta band). Preliminary results with four participants suggestthat BCI inhibitor system can improve BCI performance.
There are several ways to improve performance of Electroencephalography (EEG) based Brain-Computer Interface (BCI) systems: improving classification methods, developing new signal pro-cessing algorithms or increasing EEG hardware efficiency are the most common directions taken [1].One recent approach intends to combine different paradigms into a hybrid system [2]. For exam-ple, the Error Related Potential can be detected during a BCI-based interaction to correct theinteraction and increase overall performance [3].However relatively few studies try to improve the performance of BCI systems by focusingon interaction techniques and usage. An approach is to use a brain switch [2, 4] which can bedescribed as a preliminary BCI that allows the user to activate the BCI interaction at will. Thissystem allows reducing false positive in asynchronous BCI context as confirmed in [5].In this paper we propose an implicit and complementary system which we have named BCIinhibitor. The main objective is to increase performance of a BCI by using a paradigm that willactivate the BCI system only if the best conditions are met, with respect to the current state ofthe user. Thus BCI inhibitor will avoid classifying EEG features when the system detects thatthe current user’s state will undoubtedly lead to an erroneous result. We evaluated our solutionthrough a pilot study involving motor imagery tasks.
BCI inhibitor can be defined as a system that pauses the BCI until specific conditions (e.g.optimum condition in terms of user’s state) are met. BCI inhibitor is in relation with the recentidea of a brain switch, but in the case of BCI inhibitor it is not the user that will intend to explicitlyactivate the BCI. Instead the BCI inhibitor monitors user state to assess the readiness of the user.For this reason BCI inhibitor can be viewed as an implicit (or passive) counterpart to the brain-switch. We expect that the combination of a classical BCI with a BCI inhibitor will result in animprovement of the overall performance, along with more comfort for the user. Recently, Panicker1 a r X i v : . [ c s . H C ] N ov t al. described a P300 speller BCI enhanced with a constant flickering [6]. The P300-based BCIwas paused when no Steady State Visual Evoked Potential (SSVEP) response was detected. Thissystem can be seen as a BCI inhibitor in which the inhibition condition is “user is not looking atthe screen” and the inhibition signal is a SSVEP measured by EEG. Several other sensor channelslike electromyography (EMG) or electrooculography (EOG) could be chosen as inhibitor signal.However the monitoring of the brain activity through EEG seems to be particularly adapted andcould provide relevant information to the BCI inhibitor system. Numerous EEG markers appearto be useful to inhibit a BCI: Error potentials, rhythms associated to attention level etc. Anotherapproach is to use features directly correlated to the BCI control signal. This will allow inhibitingthe system until some specific conditions about the control signal are met. The usage of BCIinhibitor seems to be relevant to both asynchronous and synchronous BCI. In the asynchronouscontext, we can for example think of a hybrid BCI consisted of a Brain switch, followed by aBCI inhibitor that ensures that the user is in the desired state before enabling the control BCIthat actually drives the interaction. In the context of synchronous BCI an inhibitor can be usedbetween phases to put the system in standby until specific conditions are satisfied.To evaluate the inhibitor process based on this last idea, we designed a hybrid BCI by combininga synchronous motor imagery based BCI that uses beta ERS posterior to feet movement (“betarebound”), with a BCI inhibitor that checks whether the signal in the beta frequency band doesnot show any burst of activity before starting the control period. Participants:
Four participants took part in the experiment, respectively aged of 27, 24, 28 and26. They had never used any motor imagery based BCI system before.
Setup:
EEG signals were recorded using a g.USBAmp (G.Tec) amplifier, sampled at 512 Hz.The setup was made of 7 electrodes, positioned according to the 10-20 system: a ground electrode(located on AFz position) and a reference electrode (located on the left earlobe), along with fivemeasurement electrodes on Cz, C1, C2, FCz and Cpz. This EEG setup allowed us to record theEEG activity related to motor imagery of the feet [7]. The application is based on the VirtualReality application “Use the force” presented by Lotte et al. [7]. A virtual spaceship is displayedon the screen. The goal is to lift the spaceship by doing motor-related tasks: real and imaginaryfeet movements. Whenever a burst is detected in the beta activity related to feet movement,the application raises the spaceship proportionally. Instructions can be displayed, asking theparticipant to either stand still, move or stop.
Procedure:
The experiment was divided into 2 parts: a baseline and the series of motor imagerytrials, using real movements or imaginary movements. The baseline consisted of a 25 sec period,where the participants were asked to stand still and relaxed, eyes opened. No feedback wasprovided during the baseline which was done once, at startup. The trial sequence is inspired bythe startup sequence of an athletics run: Ready, Steady, Go. During one trial, the participantswere instructed to relax (“Ready”) for a certain period of time, then waited for 1 sec (“Steady”).Finally the “Move” instruction is displayed during 3 sec, followed by “Stop” during 3 sec. Themovement of the feet done during the “Move” phase was instructed to be either real or imaginary.The “Stop” instructed the user to stop doing movement which should induced a beta rebound.The BCI inhibitor was either activated or deactivated during these trials without telling theparticipants. The “Ready” phase lasted 3 sec without inhibitor. Once the inhibitor was activated,the duration of the “Ready” phase varied from a minimum of 0.5 sec to a maximum of 10 sec.Participants were asked to start with 6 real movement sessions, followed by 6 imaginary movementsessions. The BCI inhibitor was activated on half of the sessions, randomized to eliminate an ordereffect. Each session was made of 10 trials with 4 sec between trials. The whole experiment (setup,trials, and questionnaire) lasted about 1 hour. 2 ignal Processing:
EEG acquisition and online processing were conducted using the open-source software OpenViBE [8]. The EEG signal is band-pass filtered in 2-40 Hz band. Then,a Laplacian spatial filter centered on Cz is computed. The signal was then filtered in the Betaband (16-24Hz). A band power technique was applied to compute the power of the Beta band.We distinguish then two signals processed: the Control Signal (CS) that is used to control thespaceship, and the Inhibitor Signal (IS) used by the BCI inhibitor to decide to either launch ornot launch the BCI-based interaction.We define the
Control Signal (CS) as the beta band power extracted on a 1 s windowevery 100 ms. The last 4 features were averaged with a moving window to produce a smoothcontrol signal. To detect the post movement Beta ERS, the CS was compared to a threshold:
Th1 = baseline mean + 3 ∗ baseline std ; where baseline mean and baseline std correspond respectively tothe average and standard deviation of the control signal during the baseline phase.We define the Inhibitor Signal (IS) as the beta band power extracted on a 2 s windowevery 500 ms. The IS was compared to a threshold
Th2 = baseline mean + 1 ∗ baseline std . If thecomputed control signal stayed below Th2
99% of the time then the inhibition was deactivatedand the BCI started. The maximum time of the inhibition was 10 s, after this the BCI startedanyway.
Table 1 shows the participants’ performance for each condition (inhibitor on vs. inhibitor off,real movements vs. imaginary movements) and the duration of the “Ready” phase . To assess theparticipants’ performance we counted the number of false positives (FP) and the number of truepositives (TP). A false positive occurs if the value of CS went at least once above Th1 during a“Move” phase. A true positive occurs if CS signal went at least once above Th1 during a “Stop”phase. What happened during the other phases was not taken into account. We also computedthe Hit-False (HF) difference which is equal to the number of TP minus the number of FP.Table 1: Performance achieved with the motor imagery BCI with and without BCI inhibitor inreal and imaginary condition and mean duration of the “Ready” phase . Last row provides theaverage values over participants.
Task Inhibitor Duration of the FP TP HF “Ready Phase” (sec)
Subject 1 Real on 1 . ± .
03 9/30 29/30 20off 3 . ± .
00 3/30 30/30 27Imaginary on 1 . ± .
63 6/30 14/30 8off 3 . ± .
00 2/30 12/30 10Subject 2 Real on 3 . ± .
94 19/30 28/30 9off 3 . ± .
00 24/30 29/30 5Imaginary on 1 . ± .
65 15/30 18/30 3off 3 . ± .
00 18/30 14/30 -4Subject 3 Real on 6 . ± .
26 14/30 30/30 16off 3 . ± .
00 21/30 30/30 9Imaginary on 4 . ± .
18 8/30 17/30 9off 3 . ± .
00 19/30 22/30 3Subject 4 Real on 3 . ± .
88 8/30 30/30 22off 3 . ± .
00 13/30 28/30 15Imaginary on 2 . ± .
86 8/30 16/30 8off 3 . ± .
00 16/30 14/30 -2Average Real on 3.84 12.5/30 29.25/30 16.75off 3 .
00 15.25/30 29.25/30 14.00Imaginary on 2.39 9.25/30 16.25/30 7.0off 3 .
00 13.75/30 15.5/30 1.75 Discussion
Results suggest that BCI inhibitor works and provides an effect on the BCI behavior. It ismaterialized by different inhibition times for each user. Even if we have to take some cautionsconsidering the limited number of participants it seems that the BCI inhibitor is able to improvethe system performance: the average Hit-False difference over subjects is higher when the inhibitorwas enabled for real movement condition (16.75 vs. 14.0) and imaginary condition (7.0 vs. 1.75).This result is mainly due to the reduction of false positive (e.g. subjects 2, 3 and 4).After each session, participants were asked whether they felt a difference between the twoconditions (with and without inhibitor) and to quantify it on a scale between 1 (not at all) and 7(very). Suprisingly the participants reported only a small difference (2 . ± . In this paper we have introduced the concept of BCI inhibitor which can be defined as a systemthat pauses the BCI until some specific conditions are met. We presented a pilot study in the caseof a synchronous motor imagery based BCI. Preliminary results with four participants suggestthat inhibition process can be used to improve system performance. These hypotheses should beconfirmed with more subjects. Future work should also address the use of BCI inhibitor for otherparadigms such as P300 and SSVEP. Exploring adapted inhibitor signals (e.g. EEG markerscorrelated to level of attention) seems to be particularly relevant in these cases.
Acknowledgments
This work was supported by the French National Research Agency within the OpenViBE2 project(ANR-09-CORD-017). The authors would also like to thank Fabien Lotte(INRIA Bordeaux Sud-Ouest) for its helpful remarks.
References [1] F. Lotte, M. Congedo, A. L´ecuyer, F. Lamarche, and B. Arnaldi. A review of classification algorithmsfor EEG-based brain–computer interfaces.
Journal of Neural Engineering , 4, 2007.[2] G. Pfurtscheller et al. The hybrid bci.
Front. Neurosci. , 2010.[3] P. W. Ferrez and J. R. Mill´an. Simultaneous real-time detection of motor imagery and error-relatedpotentials for improved BCI accuracy. In , 2008.[4] S.G. Mason and G.E. Birch. A brain-controlled switch for asynchronous control applications.
Biomed-ical Engineering, IEEE Transactions on , 47(10):1297–1307, oct 2000.[5] G. Pfurtscheller, T. Solis-Escalante, R. Ortner, P. Linortner, and G.R. Muller-Putz. Self-paced opera-tion of an ssvep-based orthosis with and without an imagery-based “brain switch”: A feasibility studytowards a hybrid bci.
IEEE Trans. Neural Syst. Rehab. , pages 409–414, aug. 2010.[6] R.C. Panicker, S. Puthusserypady, and Ying Sun. An asynchronous p300 bci with ssvep-based controlstate detection.
Biomedical Engineering, IEEE Transactions on , 58(6):1781–1788, june 2011.[7] F. Lotte, Y. Renard, and A. L´ecuyer. Self-Paced Brain-Computer Interaction with Virtual Worlds: AQuantitative and Qualitative Study “Out of the Lab”. In , 2008.[8] Y. Renard et al. OpenViBE: An Open-Source Software Platform to Design, Test and Use Brain-Computer Interfaces in Real and Virtual Environments.
Presence , 19:35–53, Apr. 2010., 19:35–53, Apr. 2010.