Development of a Human Machine Interface for Control of Robotic Wheelchair and Smart Environment

Transcription

1 Preprints of the 11th IFAC Symposium on Robot Control, Salvador, BA, Brazil, August 26-28, 2015 ThMS2.1 Development of a Human Machine Interface for Control of Robotic Wheelchair and Smart Environment Richard J. M. G. Tello, Alexandre L. C. Bissoli, Flavio Ferrara, Sandra Müller, Andre Ferreira, Teodiano F. Bastos-Filho Post-Graduate Program in Electrical Engineering (PPGEE). Federal University of Espirito Santo (UFES). Av. Fernando Ferrari 514. Vitoria, Brazil. ( richard@ele.ufes.br; alexandre-bissoli@hotmail.com; andrefer@ele.ufes.br; teodiano.bastos@ufes.br) Politecnico di Milano: Piazza Leonardo da Vinci, 20133, Milano, Italy ( femferrara@gmail) Electrical Engineering Department, Federal Institute of Esprito Santo (IFES). Av. Vitoria, 1729, Vitoria, Brazil ( sandra.muller@ifes.edu.br) Abstract: In this work, we address the problem of integrating a robotic wheelchair into a smart environment. This approach allows people with disabilities to control home appliances of the environment using a Human Computer Interface (HCI) based on different biological signals. The home appliances includes TV, radio, lights/lamp and fan. Three control paradigms using surface Electromyography (semg), Electrooculography (EOG) and Electroencephalography (EEG) signals were used. These signals are captured through a biosignal acquisition system. Three subparadigms for semg/eog analyzes were defined: moving eyes horizontally (left/right), raising brow and prolonged clench. On the other hand, the navigation of the wheelchair is executed through an Steady-State Visually Evoked Potentials (SSVEP)-BCI. Each stage of our proposed system showed a good performance for most subjects. Therefore, volunteers were recruited to participate of the study and were distributed in two groups (subjects for home appliances and subjects for SSVEP-BCI). The average accuracy for prolonged clench approach was of 95%, the raising brow was 85% and moving eyes achieved 93%. Multivariate Synchronization Index (MSI) was used for feature extraction from EEG signals. The flickering frequencies were 8.0 Hz (top), 11.0 Hz (right), 13.0 Hz (bottom) and 15.0 Hz (left). Results from this approach showed that classification varies in the range of 45-77% among subjects using window length of 1 s. Keywords: SSVEP-BCI, Robotic Wheelchair, EEG, semg, EOG, Smart Environment. 1. INTRODUCTION A Human Machine Interface (HMI) is a platform that allows interaction between user and automatized system. On the other hand, a Brain-computer interface (BCI) is a technology that provides human with direct communication between the users brain signals and a computer, generating an alternative channel of communication that does not involve the traditional way as muscles and nerves (Wolpaw et al. (2000)). Among current BCIs, a noninvasive brain imaging method commonly employed in BCIs is EEG, which has the advantages of lower risk, inexpensive and easily measurable (Chen et al. (2014) and Kelly et al. (2005)). Further, EEG provides electrical signals of high The authors thank FAPES (a foundation of the Secretary of Science and Technology of the State of Espirito Santo, Brazil), CAPES (a foundation of the Brazilian Ministry of Education) and CNPQ (The Brazilian National Council for Scientific and Technological Development), for the support given to this work. temporal resolution generated by neuronal dynamics from the scalp. Therefore, a BCI records brain signals, and EEG signal features are then translated into artificial outputs or commands that act in a real world. BCI is a potential alternative and augmentative communication (AAC) and control solution for people with severe motor disabilities (Wolpaw et al. (2000), Kelly et al. (2005) and Gao et al. (2003)). One kind of BCI named SSVEP-BCI uses the excitation of the retina of eye by a stimulus at a certain frequency, making the brain generating an electrical activity of the same frequency with its multiples or harmonics. This stimulus produces a stable Visual Evoked Potential (VEP) of small amplitude termed as Steady-State Visually Evoked Potentials (SSVEPs) of the human visual system. To produce such potentials, the user gazes at one flickering stimulus oscillating at a certain frequency (He (2013)). In a typical SSVEP-BCI system, several stimuli flickering at different frequencies are presented to the user. The subject Copyright 2015 IFAC. 144

2 overtly directs attention to one of the stimuli by changing his/her gaze attention (Zhang et al. (2010)). This kind of SSVEP-BCI was evaluated in this study and is commonly called as dependent since muscle activities, such as gaze shifting, are necessary. One of the first studies related to control of smart home applications using biological signals, such as EEG, was reported in (Holzner et al. (2009)). In that work a BCI based on P300 approach is used for TV channels switching, for opening and closing doors and windows, navigation and conversation, but all in a controlled environment of a virtual reality (VR) system. Twelve subjects were evaluated and an average of 67.51% in the classification for all subjects and all decisions was achieved. Other study (Ou et al. (2012)) based on VR in order to create a controlled environment was performed. In that work, the term Brain computer interface-based Smart Environmental Control System (BSECS) was introduced and a BCI technique with Universal Plug and Play (UPnP) home networking for smart house applications environment was proposed. Also, an architecture where the air conditioner and lights/lamp can be successfully and automatically adjusted in realtime based on the change of cognitive state of users was designed. A hybrid BCI for improving the usability of a smart home control was reported in (Edlinger and Guger (2012)). In that study, P300 and SSVEP approaches were used. Results indicated that P300 is very suitable for applications with several controllable devices, where a discrete control command is desired. However, that study also reports that SSVEP is more suitable if a continuous control signal is needed and the number of commands is rather limited. A simple threshold criterion was used to determine if the user is looking at the flickering light. All the different commands were summarized in 7 control masks: a light mask, a music mask, a phone mask, a temperature mask, a TV mask, a move mask and a go to mask. That study was also tested in a VR. A similar approach using a hybrid BCI paradigm based on P300 and SSVEP is reported in (Wang et al. (2014)), where a Canonical Correlation Analysis (CCA) technique was applied for the SSVEP detection. Applications involving robotic wheelchairs and SSVEP signals were also reported in (Muller et al. (2011); Xu et al. (2012); Diez et al. (2013); Singla et al. (2014)). A recent study using SSVEP and P300 approaches for wheelchair control was reported in (Li et al. (2013)). On the other hand, a hybrid BCI based on SSVEP and visual motion stimulus was applied to a robotic wheelchair in (Punsawad and Wongsawat (2013)). Finally, studies that combine motor imagery (MI) and SSVEPs to control a real wheelchair were reported in (Bastos et al. (2011) and Cao et al. (2014)). In this work, we address the problem of integrating a wheelchair into a smart environment. Due to the variety of disabilities that benefit from assistive technologies, an optimal approach could allow the user to choose the preferred control paradigm according to the degree of his/her disability. The system allows the handling of various devices in a real environment, e.g. a room, by means of biological signals controlled from a robotic wheelchair. We present this system with three kind of assistive control paradigms using, respectively, muscle (semg), EOG and brain signals (EEG) as shown in Fig. 1. USER INTERFACE semg SIGNAL EOG SIGNAL EEG SIGNAL LEVELS OF CAPACITY Face muscles: clenching and raising brow Ocular globe: left and right SSVEP Fig. 1. Levels of capacity of our proposed system. 2. METHODS Different stages were addressed as follows: 2.1 Assistive System In the context of smart environments applied to assistive technologies, this paper proposes an input interface allowing people with disabilities to turn on and off appliances without help, from the wheelchair. Part of this work was to design and build a smart box that allows controlling up to four appliances in an environment, including TV set, radio, lamp/lights and fan. semg, EOG and EEG signals were recorded using a biological signal acquisition device. The user can issue commands from the wheelchair, and then the signal is transmitted through RF to the smart box, when the corresponding equipment is finally operated. We used RF communication to turn the appliances on and off remotely. The RF transmitter and receiver work in a frequency of 433MHz, controlled by an Arduino Mega microcontroller. The communication is unidirectional, that is, only the transmitter sends the data to the receiver. On the other hand, an SSVEP-BCI was used to control the navigation of the wheelchair. The wheelchair is equipped with a small 10.1 display that exhibits the Control Interface (CI). In addition, four small boxes containing four Light-Emitting Diodes (LEDs) of white color were placed in the four side of this display as visual stimuli for the generation of evoked potentials. The CI uses the display to visualize a menu, through which the user can navigate or operate the desired device. It is worth noting that this menu is dynamic, as device options can change according to the current room, or be customized by the user before running the system. Moreover, for some device we provide additional operations that can be performed using a sub-menu. For example, after turning on the TV, the display shows a sub-menu with options such as Channel Up, Channel Down, Volume Up, Volume Down. It is always possible to go back to the main menu and turn off the system. The CI was presented in (Ferrara et al. (2015)). It offers an interface of procedures that can be accessed through Remote Procedure Call (RPC), that is, one for turning the interface on, one for turning it off, and one for transmitting Copyright 2015 IFAC. 145

3 logical commands encoded as integers. In (Ferrara et al. (2015)), it has been demonstrated that choosing three kinds of paradigms is optimal to operate the menu. In this work, we used a novel method to assess the overall performance of the system, named Utility (Dal Seno et al. (2010)) while using the preferred control paradigm. 2.2 semg and EOG control For individuals with disabilities that do not affect voluntary control of facial muscles, we propose a system based on semg and EOG signals. With the aim of ensuring a high reliability, we defined the three paradigms aforementioned: moving eyes horizontally (left/right), raising brow and prolonged clench. For processing semg/eog signals, a comparison of the signal amplitude with a predefined threshold value was performed. Aspects involving the width and duration of the signal components in the decision of classification were observed. The signals corresponding to the left-right eye movements were dominant in the horizontal axes. For those signals, we found an amplitude and duration very prolonged, opposite signals and high amplitude. Thus, the user is able to control the options through moving eyes whereas the raising brow is used to confirm the highlighted option. Finally, prolonged clench option is used for activation and deactivation of the SSVEP-BCI. Fig. 2 shows a summary of the three paradigms allocated for the semg/eog approaches. Fig. 2. Biological signal transducer module for semg/eog signals. Each individual performed the tasks regarding the semg/ EOG in a different personal fashion. Thus, it is quite optimistic to expect that a new user is able to achieve an optimal performance since the first trial. However, our tests revealed than for most users, just a very brief adaptation period was required to figure out the best way to perform the gesture. This adaptation period consisted of 2 or 3 minutes while the user tries to execute a command and observe a visual feedback in a LCD screen. Although this step is not required, it is recommended considering that this procedure can accelerate the development of the user s control skills. Online experiments with eight healthy subjects were performed. Further, the subjects were seated on the wheelchair, in front of the display where the program was running, and asked to perform thirty repetitions per command, resulting in 90 total trials. In a control panel, the level of clenching and raising brow through continuous values between 0 and 1 is expressed, so it is sufficient to set predefined thresholds to detect the two movements. We opted to trigger an action when the correspondent value is greater than 0.8 and the other one is less than 0.2. These thresholds were obtained empirically; we found them to provide good results with most subjects. For detecting eye movements, our system provides a discrete number with values 0 and 1. We analyzed the classification accuracy of 720 total trials. Hence, the average accuracy and trial duration using Utility were computed. 2.3 Estimation of MSI for SSVEP-BCI This stage is in charge of controlling the navigation of the robotic wheelchair. Five subjects (three males and two females), with ages from 21 to 27 years old, were recruited to participate in this study (average age: 25.6; Standard Deviation (STD): 2.61). The research was carried out in compliance with Helsinki declaration, and the experiments were performed according to the rules of the ethics committee of UFES/Brazil, under registration number CEP- 048/08. For the development of the SSVEP-BCI, 12 channels of EEG with the reference at the left ear lobe were recorded at 600 samples/s, with 1 to 100 Hz pass-band. The ground electrode was placed on the forehead. The EEG electrode placements were based on the International System. The electrodes used were: P7, PO7, PO5, PO3, POz, PO4, PO6, PO8, P8, O1, O2 and Oz. The equipment used for EEG signal recording was BrainNet-36. The timing of the four LEDs flickers was precisely controlled by a microcontroller (PIC18F4550, Microchip Technology Inc., USA) with 50/50% on-off duties, and frequencies of 8.0 Hz (top), 11.0 Hz (right), 13.0 Hz (bottom) and 15.0 Hz (left). To send commands to the wheelchair, the user has to fix the attention to one of the flickering frequencies. The EEG data are segmented and windowed in window lengths (WL) of 1 s with an overlap of 50%. Then, a spatial filtering is applied using a Common Average Reference (CAR) filter and a band-pass filter between 3-60 Hz for the twelve channels. Several studies (Vialatte et al. (2010); Pastor et al. (2003)) confirm that visual evoked potentials are generated with greater intensity on the occipital area of the cortex. Thus, the twelve electrodes were used in the initial stage only for application of a CAR spatial filter. According to our observations, the application of this spatial filter to the twelve electrodes improves the classification performance when selecting O1, O2 and Oz electrodes. Based on that fact, we have evaluated the detection of SSVEPs using these three channels as input vector for the feature extractor after the filtering process. Multivariate Synchronization Index (MSI) was used for feature extraction. A brief description of this technique is explained below. MSI is a novel method to estimate the synchronization between the actual mixed signals and the reference signals as a potential index for recognizing the stimulus frequency. (Zhang et al. (2014)) has proposed the use of a S-estimator as index, which is based on the entropy of the normalized eigenvalues of the correlation matrix of multivariate signals. Autocorrelation matrices C 11 and C 22 for X and Y i, respectively, and crosscorrelation matrices C 12 and C 21 for X and Y i can be obtained as (Tello et al. (2014)), where i refers to the number of targets Copyright 2015 IFAC. 146

4 C 11 = (1/N).XX T (1) T C 22 = (1/N).Y i.y i (2) T C 12 = (1/N).XY i (3) C 21 = (1/N).Y i.x T (4) A correlation matrix C i can [ be constructed ] as C i C11 C = 12 C 21 C 22 (5) in the lateral canthus of the eye (F7 and F8 position according to standard) in order to monitoring the frontal lobe. In addition, Fig. 4 shows the general block diagram of the proposed system. The internal correlation structure of X and Y i contained in the matrices C 11 and C 22, respectively, is irrelevant for the detection of stimulus frequency (Carmeli et al. (2005)). It can be removed by constructing a linear transformation matrix U = [ C11 1/2 0 0 C 22 1/2 ] (6) So that C 1/2 11 C 1/2 11 =C 11, C 1/2 22 C 1/2 22 =C 22 and by applying the transformation C i = UCU which results in a transformed correlation matrix of size P P, where P = M + 2H (Carmeli et al. (2005)). The eigenvalues λ i 1, λ i 2,...,λ i P of C i, normalized as λ i m = λ i m/ P m=1 λi m for m = 1, 2,..., P, can be used to evaluate the synchronization index S i for matrix Y i as P m=1 S i = 1 + λ i m log( λ i m) (7) log(p ) see (Zhang et al. (2014)). Using S 1,S 2,...,S K computed for the stimulus frequencies f 1,f 2,...,f K, the MSI can be estimated as S = maximum S i (8) 1 i K In this case, the way to assess the performance of the SSVEP-BCI system was the Shannon s Information Transfer Rate (ITR), see details in (Vialatte et al. (2010)). 3. SYSTEM ARCHITECTURE semg, EOG and EEG signals are captured by the signal acquisition equipment, which has inputs for electromyographic and electroencephalographic signals. Through a computational sniffer the biological signals are read from the equipment, then these signals are transmitted and processed in an embedded computer by algorithms developed in Matlab. This embedded computer in the wheelchair has the following specifications: Mini ITX motherboard, 3.40 GHz Intel Core i5 processor, and 4GB RAM. The data are analyzed in the main routine and the prolonged clench signal works as a switch, which determines the operation of the navigation or the smart environment control. Raising brow or moving the eyes are used to control the devices inside the house. The navigation of the wheelchair is executed through the SSVEP-BCI approach, which is based on commands and these are used for directional control of the wheelchair. The LED on the above side (8 Hz) indicates forward, the LED on the right indicates the movement to the right turn (11 Hz), the LED on the left side (15 Hz) indicates the movement to the left turn and, finally, the LED on the bottom (13 Hz) indicates stopping. Fig. 3 shows all electrode placement locations and a user on the wheelchair. The semg/eog electrodes were placed Fig. 3. (a) Electrode placement location using system for our system; (b) an user using the wheelchair. 4. EXPERIMENTAL RESULTS Table 1 summarizes the classification outcome for muscle movement tasks. Each movement is treated independently because a subject may be able to execute some movement very better (or worse) than the others. It can be noticed that the average accuracy is remarkably higher than random guessing and often expresses a very good performance of the classifier (between 85% to 95%). Nevertheless, as often in assistive technologies, some users show troubles while realizing a certain movement that preclude them to give the correspondent command. For example, subject 1 was not able to succeed while raising brow. Since the computation of Utility depends on which menu option the user wants to operate, we consider the mean Utility as the arithmetic mean of the Utility in a menu with four options and express the result in bits per minute to facilitate the comparison with ITR. A critical advantage of control by means of muscle movement is the speed of recognition. It can be seen that an operation is triggered after a very small amount of time. Fig. 5 represents the results of online tests using facial expressions. Each point represents an expression made by a user. Points in the top-left corner indicate better performance. The size of the dots represents the value of Utility. This could lead to a fast and efficient control paradigm, especially after a period of training and selfimprovement. For evaluation of the EEG control, each volunteer (during 30 seconds) fixes his/her attention to each stimulus and the results are shown in Table 2. The results from the SSVEP- BCI were acceptable considering that subjects never used a BCI and neither had previous training. The highest value of accuracy was for the subject 3 with 77% and bits/min of ITR achieved from the average value of the accuracy. Copyright 2015 IFAC. 147

5 Navigation Control Pre-processing Feature Extraction Classification semg (Clenching) Main Routine PC Switch Control Brain Signals (EEG) Biosignal Acquisition Equipment EOG + semg (raising brow) Smart Environment UFES Robotic Wheelchair Brain Signals (EEG) Ocular globe (EOG) Switch Control 8 Hz Face muscles (semg) 15 Hz 13 Hz 11 Hz Visual stimuli through SSVEP Navigation Fig. 4. The system paradigm of our multimodal system. Fig. 5. Comparison between control of the system from different muscle movements and speed of recognition. Table 1. Accuracy results for EEG control using WL of 1 s. SSVEP Subjects Frequency s1 s2 s3 s4 s5 8 Hz Hz Hz Hz Mean Acc. [%] ITR [bits/min] CONCLUSION In this paper we presented a multimodal system capable to employ different biological signals in order to control several home appliances and the navigation of a robotic wheelchair. Three control paradigms using semg, EOG and EEG signals were introduced and each stage showed a good performance for most subjects. Our strategy provides the reliability in terms of classification result and safety of wheelchair control. To evaluate the semg/eog system, eight subjects participated of the experiments. The average accuracy for prolonged clench approach was of 95%, the raising brow was 85% and moving eyes achieved 93%. On the other hand, five subjects participated of the study for the SSVEP-BCI. The total values of classification vary in the range of 45-77% among subjects, considering a WL of 1s. Our results would even improve because it is widely known that increase in accuracy is related with increase of the time windows (Tello et al. (2014)), this fact suggests that while more information is processed, the feature extractor can detect visual evoked potentials with more precision. ACKNOWLEDGEMENTS The authors wish to thank all the volunteers for their participation in the experiments. REFERENCES Bastos, T., Muller, S., Benevides, A., and Sarcinelli-Filho, M. (2011). Robotic wheelchair commanded by SSVEP, motor imagery and word generation. In IEEE Engineering in Medicine and Biology Society, (EMBC). Cao, L., Li, J., Ji, H., and Jiang, C. (2014). A hybrid brain computer interface system based on the neurophysiological protocol and brain-actuated switch for wheelchair control. Journal of Neuroscience Methods, 229, Carmeli, C., Knyazeva, M.G., Innocenti, G.M., and Feo, O.D. (2005). Assessment of EEG synchronization based on state-space analysis. NeuroImage, 25, Copyright 2015 IFAC. 148

6 Chen, C.H., Ho, M.S., Shyu, K.K., Hsu, K.C., Wang, K.W., and Lee, P.L. (2014). A noninvasive brain computer interface using visually-induced near-infrared spectroscopy responses. Neur. Letters, 580, Dal Seno, B., Matteucci, M., and Mainardi, L. (2010). The Utility Metric: A Novel Method to Assess the Overall Performance of Discrete Brain-Computer Interfaces. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 18(1), Diez, P.F., Muller, S.M.T., Mut, V.A., Laciar, E., Avila, E., Bastos-Filho, T.F., and Sarcinelli-Filho, M. (2013). Commanding a robotic wheelchair with a highfrequency steady-state visual evoked potential based brain-computer interface. Med Eng Phys. Edlinger, G. and Guger, C. (2012). A hybrid Brain- Computer Interface for improving the usability of a smart home control. In Complex Medical Engineering (CME), 2012 ICME, Ferrara, F., Bissoli, A., and Bastos-Filho, T. (2015). Designing an Assistive Control Interface based on Utility. Proceedings of the 1st International Workshop on Assistive Technology IWAT 2015, Vitoria, Brazil, Gao, X., Xu, D., Cheng, M., and Gao, S. (2003). A BCIbased environmental controller for the motion-disabled. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2), He, B. (2013). Neural Engineering. Springer. 2nd ed. Holzner, C., Guger, C., Edlinger, G., Gronegress, C., and Slater, M. (2009). Virtual Smart Home Controlled by Thoughts. In Enabling Technologies: Infrastructures for Collaborative Enterprises, WETICE th IEEE International Workshops on, Kelly, S., Lalor, E., Reilly, R., and Foxe, J. (2005). Visual spatial attention tracking using high-density SSVEP data for independent brain-computer communication. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 13(2), Li, Y., Pan, J., Wang, F., and Yu, Z. (2013). A Hybrid BCI System Combining P300 and SSVEP and Its Application to Wheelchair Control. Biomedical Engineering, IEEE Transactions on, 60(11), Muller, S.M.T., de S, A.M.F.L.M., Bastos-Filho, T.F., and Sarcinelli-Filho, M. (2011). Spectral techniques for incremental SSVEP analysis applied to a BCI implementation. V CLAIB La Habana, Ou, C.Z., Lin, B.S., Chang, C.J., and Lin, C.T. (2012). Brain Computer Interface-based Smart Environmental Control System. In Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), Pastor, M., Artieda, J., Arbizu, J., Valencia, M., and Masdeu, J. (2003). Human cerebral activation during steady-state visual evoked response. J. Neurosci. Punsawad, Y. and Wongsawat, Y. (2013). Hybrid SSVEPmotion visual stimulus based BCI system for intelligent wheelchair. In Engineering in Medicine and Biology Society (EMBC), Singla, R., Khosla, A., and Jha, R. (2014). Influence of stimuli colour in SSVEP-based BCI wheelchair control using support vector machines. Journal of Medical Engineering & Technology, 38, Tello, R., Muller, S., Bastos-Filho, T., and Ferreira, A. (2014). A comparison of techniques and technologies for SSVEP classification. In Biosignals and Biorobotics Conference (BRC), 5th ISSNIP-IEEE, 1 6. Vialatte, F.B., Mauriceb, M., Dauwelsc, J., and Cichocki, A. (2010). Steady state visually evoked potentials: Focus on essential paradigms and future perspectives. Progress in Neurobiology, 90, Wang, M., Daly, I., Allison, B.Z., Jin, J., Zhang, Y., Chen, L., and Wang, X. (2014). A new hybrid BCI paradigm based on P300 and SSVEP. Jour of Neur Methods. Wolpaw, J., Birbaumer, N., Heetderks, W., McFarland, D., Peckham, P., Schalk, G., Donchin, E., Quatrano, L., Robinson, C., and Vaughan, T. (2000). Brain-computer interface technology: a review of the first international meeting. Rehabilitation Engineering, IEEE Transactions on, 8(2), Xu, Z., Li, J., Gu, R., and Xia, B. (2012). Steady- State Visually Evoked Potential (SSVEP)-Based Brain- Computer Interface (BCI): A Low-Delayed Asynchronous Wheelchair Control System. Neural Information Processing. 19th ICONIP. Springer., Zhang, D., Maye, A., Gao, X., Hong, B., Engel, A.K., and Gao, S. (2010). An independent brain-computer interface using covert non-spatial visual selective attention. J. Neural Eng. Zhang, Y., Xu, P., Cheng, K., and Yao, D. (2014). Multivariate synchronization index for frequency recognition of SSVEP-based brain computer interface. Journal of Neuroscience Methods, 221(0), Table 2. Summary results for semg/eog control. Prolonged clench Subject Acc. [%] Time [s] Utility [bits/min] Average Raising brow Subject Acc. [%] Time [s] Utility [bits/min] Average Moving eyes Subject Acc. [%] Time [s] Utility [bits/min] Average Copyright 2015 IFAC. 149