Home

Search

Collections

Journals

About

Contact us

My IOPscience

A brain–computer interface based on self-regulation of gamma-oscillations in the superior parietal cortex

This content has been downloaded from IOPscience. Please scroll down to see the full text. 2014 J. Neural Eng. 11 056015 (http://iopscience.iop.org/1741-2552/11/5/056015) View the table of contents for this issue, or go to the journal homepage for more Download details: IP Address: 207.162.240.147 This content was downloaded on 14/06/2017 at 12:11 Please note that terms and conditions apply.

You may also be interested in: Self-regulation of brain rhythms in the precuneus: a novel BCI paradigm for patients with ALS Tatiana Fomina, Gabriele Lohmann, Michael Erb et al. High gamma-power predicts performance in sensorimotor-rhythm brain--computer interfaces Moritz Grosse-Wentrup and Bernhard Schölkopf An independent brain--computer interface using covert non-spatial visual selective attention Dan Zhang, Alexander Maye, Xiaorong Gao et al. Achieving a hybrid brain–computer interface with tactile selective attention and motor imagery Sangtae Ahn, Minkyu Ahn, Hohyun Cho et al. Eye-gaze independent EEG-based brain--computer interfaces for communication A Riccio, D Mattia, L Simione et al. Ensembles of adaptive spatial filters increase BCI performance: an online evaluation Claudia Sannelli, Carmen Vidaurre, Klaus-Robert Müller et al. Brain–computer interfacing under distraction: an evaluation study Stephanie Brandl, Laura Frølich, Johannes Höhne et al. An independent SSVEP-based brain--computer interface in locked-in syndrome D Lesenfants, D Habbal, Z Lugo et al. Leveraging anatomical information to improve transfer learning in brain–computer interfaces Mark Wronkiewicz, Eric Larson and Adrian K C Lee

Journal of Neural Engineering J. Neural Eng. 11 (2014) 056015 (13pp)

doi:10.1088/1741-2560/11/5/056015

A brain–computer interface based on selfregulation of gamma-oscillations in the superior parietal cortex Moritz Grosse-Wentrup and Bernhard Schölkopf Department Empirical Inference, Max Planck Institute for Intelligent Systems, Spemannstrasse 38, D-72076 Tübingen, Germany E-mail: [email protected] and [email protected] Received 22 January 2014, revised 14 July 2014 Accepted for publication 18 July 2014 Published 15 August 2014 Abstract

Objective. Brain–computer interface (BCI) systems are often based on motor- and/or sensory processes that are known to be impaired in late stages of amyotrophic lateral sclerosis (ALS). We propose a novel BCI designed for patients in late stages of ALS that only requires high-level cognitive processes to transmit information from the user to the BCI. Approach. We trained subjects via EEG-based neurofeedback to self-regulate the amplitude of gamma-oscillations in the superior parietal cortex (SPC). We argue that parietal gamma-oscillations are likely to be associated with high-level attentional processes, thereby providing a communication channel that does not rely on the integrity of sensory- and/or motor-pathways impaired in late stages of ALS. Main results. Healthy subjects quickly learned to self-regulate gamma-power in the SPC by alternating between states of focused attention and relaxed wakefulness, resulting in an average decoding accuracy of 70.2%. One locked-in ALS patient (ALS-FRS-R score of zero) achieved an average decoding accuracy significantly above chance-level though insufficient for communication (55.8%). Significance. Self-regulation of gamma-power in the SPC is a feasible paradigm for brain–computer interfacing and may be preserved in late stages of ALS. This provides a novel approach to testing whether completely locked-in ALS patients retain the capacity for goal-directed thinking. Keywords: brain–computer interface, brain–machine interface, ALS, locked-in, gammaoscillations, neurofeedback (Some figures may appear in colour only in the online journal) 1. Introduction

instance, volitional modulation of sensorimotor rhythms (SMRs) [5] is one of the most prevalent BCI paradigms [9], even though ALS is known to be associated with a degeneration of neurons in the primary motor cortex [10] and a decrease in resting-state SMRs [11]. Visual P300 speller systems [4, 12], on the other hand, require subjects to fixate target letters for optimal decoding performance [13]. This is not practicable for patients in late stages of ALS due to impaired oculomotor control [14]. To address this issue, novel BCI paradigms based on covert attention shifts have been developed. These require subjects only to fixate one spot, while covertly shifting attention to different peripheral locations of the visual field [15] or attending to various aspects,

Brain–computer interfaces (BCIs) hold the promise of enabling communication with completely locked-in (CLI) patients in late stages of amyotrophic lateral sclerosis (ALS). Despite intensive research efforts spanning four decades [1–6], this promise has yet to be fulfilled [7]. The reasons for the present inability of CLI patients to communicate with a BCI remain largely unknown. It has been argued that longterm paralysis leads to the extinction of goal-directed thinking [8], with CLI patients residing in a state-of-mind similar to REM sleep [7]. An alternative explanation is that BCIs are often based on processes known to be impaired in ALS. For 1741-2560/14/056015+13$33.00

1

© 2014 IOP Publishing Ltd Printed in the UK

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

e.g. color or form, of visual stimuli [16]. As CLI patients are by definition unable to fixate any location of the visual field, they are unlikely to benefit from these developments. Tactile P300 systems [17] may suffer from similar problems, as vibrotactile stimulation in one ALS patient in the CLI state was not found to result in any changes in cortical activity [18]. Auditory BCIs may provide a viable alternative [19–21], as processing of auditory stimuli has been found to remain intact in the CLI state [18]. Despite promising results of auditory paradigms with severely paralyzed patients [22, 23], no successful communication attempt with a CLI patient has been reported. It thus remains unclear whether CLI patients are unable to use a BCI due to the extinction of goal-directed thinking, or because available systems are based on neural and/or cognitive processes that are impaired in late stages of ALS. We argue that in order to decide this question, it is essential to develop BCI paradigms that are not based on sensory- and/or motor-processes known to be impaired in ALS. Specifically, we propose to train subjects to self-regulate spontaneous activity in high-level brain areas that are not known to be directly affected by ALS and that are likely to remain engaged during states of reduced consciousness. This approach is conceptually similar to the self-regulation of slow cortical potentials (SCPs) pioneered by Birbaumer et al in the 1990s [3, 24, 25]. However, self-regulation of SCPs requires extensive training [24] and has not been demonstrated in the CLI state [7]. Moreover, the cognitive correlates of SCPs remain debated [26, 27]. This raises the question which neural signals representing high-level cognitive processes are amenable to self-regulation by severely paralyzed patients. In this work, we investigate the utility of γ-oscillations in the superior parietal cortex (SPC) for this purpose. Oscillations in the γ-range (>40 Hz) are an indicator of local information processing by inhibitory interneurons [28–30] and are known to be involved in a variety of cognitive tasks [31, 32]. Due to their potential contamination by electromyogenic (EMG) artifacts, they have long been shunned in EEG research [33]. However, a variety of groups have demonstrated that spatial filtering allows the observation of cortical oscillations in the γ-range in EEG recordings [34–37]. We have provided empirical evidence that γ-oscillations originating in the SPC can be observed in EEG recordings, and argued that they are likely to be generated by a high-level attention network that modulates sensorimotor areas [36, 38]. As such, we consider them less likely to be affected by ALS than brain activity more closely linked to the peripheral nervous system. Furthermore, as the SPC has been found to remain engaged during REM sleep [39], we consider it a primary target for interacting with patients who may be in a state-of-mind akin to REM sleep. It is an open question, however, whether γ-oscillations in the SPC can be selfregulated and thus employed to transmit information. In this work, we show that it can based on experimental data from eleven healthy subjects and one (not yet completely) lockedin ALS patient. We emphasize that it is not our goal at this stage to design an assistive communication device with a high information transfer rate for patients with residual muscle

Figure 1. Visual feedback shown to subjects during neurofeedback

training. Subjects employed self-regulation of γSPC to move the ball vertically towards the current target, indicated by the yellow rectangle.

activity. Rather, we aim to establish a BCI paradigm for basic yes/no-communication that offers a novel approach to challenging the thought-extinction hypothesis in CLI patients in late stages of ALS.

2. Methods Sections 2.1–2.4 focus on the experimental setup and analysis procedures for the healthy subjects. Section 2.5 is devoted to the feedback sessions carried out with the locked-in ALS patient. 2.1. Experimental paradigm

Subjects were placed in a comfortable chair in a dimly lit room approximately 1.25 m away from a computer screen. Prior to each feedback session, a five-minute resting-state EEG was recorded. The subject was instructed to let their mind wander while fixating a gray cross displayed centrally on the computer screen. Data recorded during the resting-state baseline was used to calibrate a beamformer employed in the subsequent feedback session. The beamformer was aimed at a region of the SPC that we found to predict performance in an SMR-based BCI [36]. The details of the beamforming procedure are described in section 2.3. Following the resting-state baseline, subjects performed three feedback blocks with brief intermissions in between. In each block, subjects received continuous feedback on their current state of γ-power in the SPC. Specifically, log-bandpower in the range of 55–85 Hz in the SPC, as estimated by the beamformer, was linearly mapped to the vertical position of a white ball displayed on the computer screen (figure 1). Subsequently, log-bandpower of γ-oscillations in the SPC is referred to as γSPC . The vertical center of the screen was chosen to represent the mean of γSPC during the resting-state baseline. The top and bottom of the screen were chosen to represent two positive and negative standard deviations of 2

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

resting-state γSPC , respectively. The vertical movement of the white ball was constrained to the limits of the computer screen. The ballʼs horizontal position was fixed at the center of the screen. Each of the three blocks consisted of 20 trials, in which subjects had to either enhance or attenuate γSPC. Every trial began with a baseline during which a fixation cross and two gray blocks were shown at the top and bottom of the screen. After 5 s, the white ball appeared and one of the gray blocks turned yellow, instructing the subject to move the ball to this target. To reinforce successful self-regulation of γSPC , the target block turned green whenever it was touched by the ball. After 60 s, the ball disappeared and the target-color changed back to gray. This indicated the end of the trial and start of the subsequent trialʼs baseline. We chose a trial length of 60 s, as we reasoned that successful self-regulation of attentional states is likely to require more time than typically alloted to subjects in BCI paradigms. For each block, the number of trials per condition was balanced and the order of trials was pseudo-randomized. No instructions were given to the subjects in order to obtain unbiased reports on the cognitive strategies employed to self-regulate γSPC . Following each recording session, subjects were asked to note down the thoughts and feelings associated with control of the feedback signal.

Figure 2. Topography of sources in the superior parietal cortex

targeted by the beamforming procedure (see section 2.3).

compute a linearly-constrained-minimum-variance (LCMV) beamformer [40]. We chose beamforming because it has been shown to achieve decoding results competitive with other state-of-the-art spatial filtering methods for non-invasive BCIs [41]. The resting-state EEG was first bandpass-filtered between 55 and 85 Hz using a third-order Butterworth filter before computing a subject-specific spatial covariance matrix Σ ∈ N × N with N = 121 (the number of EEG channels). The beamformer was then computed by solving the LCMV-optimization problem

2.2. Experimental data

Twenty-one healthy subjects participated in this study (ten female; mean age of 31.9 years with a standard deviation of 11.4 years). Two subjects were excluded from the analysis due to medication-related issues. The remaining nineteen subjects are referred to as S1–S19. Following the initial feedback session, three subjects were re-invited to receive four additional training sessions. These subjects were chosen based on their decoding accuracy in the first recording session, with one subject performing very well in the initial session (S3), one subject showing a moderate performance (S8), and one subject having no initial control of the feedback signal (S10). The additional training sessions were carried out at the same time of the day and on the same day of the week on four consecutive weeks. Four subjects had previous experience with either motor-imagery or P300-based BCIs (S1, S8, S10 and S11). All subjects gave informed consent in agreement with guidelines of the Max Planck Society. Throughout all sessions a 121-channel EEG was recorded at a sampling frequency of 500 Hz using actiCAP active electrodes and a QuickAmp amplifier (both provided by BrainProducts GmbH, Gilching, Germany). Electrodes were placed according to the extended 10–20 system with electrode Cz as the initial reference. All recordings were converted to common average reference.

{

}

w* = argmin w TΣw s. t. w Ta = 1 w

(1)

with a ∈ N the scalp topography generated by the targeted dipoles in the SPC. This scalp topography was obtained by generating a forward model, described by a leadfield matrix L ∈ N × K for K = 15028 dipoles distributed throughout the cortex, and taking the mean of L over the subset of 300 dipoles found to be most predictive of performance in an SMR-based BCI [36]. This resulted in the topography a displayed in figure 2, exhibiting a focus over right SPC. The leadfield matrix L was generated with the BrainStorm toolbox [42], using standardized electrode locations and a standardized three-shell spherical head model. The solution to the optimization problem in (1) is given by T −1 −1 T −1 * w = a Σ a a Σ [40]. The resulting spatial filter w*

(

)

was then applied to the EEG data x [t ] ∈ N to obtain a 1Dsignal y [t ] = w*T x [t ]. In this signal, EEG sources in the targeted region are represented with unit gain, while the variance of all EEG sources outside the targeted area is optimally attenuated. In order to estimate the beamformerʼs spatial transfer function, i.e., the gain with which activity at each cortical location is passed into the spatially filtered EEG signal, the EEG data can be modeled as y [t ] = w*T x [t ] = w*T Ls [t ], with s [t ] ∈ K the (unknown) time-course of cortical

2.3. Online feedback

In order to provide each subject with online feedback on their current level of γ-power in the SPC, EEG data recorded during the subjectʼs resting-state baseline was used to 3

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

Figure 3. Group-average of the beamformerʼs spatial transfer function (gain vector g in section 2.3), with a focus in the right superior parietal

cortex.

sources. The beamformerʼs spatial transfer function is then given by g = w*T L . This gain vector can be plotted in 3D to obtain a visualization of the cortical areas used for online feedback (see figure 3 in section 3.1). For each subject, the beamformer computed from the resting-state data was held constant throughout all feedback sessions, as dynamic adaptation of beamformers has been found to have little influence on decoding performance [41]. During the feedback sessions, the current estimate of γSPC was obtained by first applying the beamformer to the recorded data and then computing log-bandpower between 55 and 85 Hz of the spatially filtered signal. Bandpower computation was based on a fast Fourier transform in conjunction with a Hanning window, using the past 5 s of recorded data. This estimate of γSPC was then standardized by the mean and standard deviation of γSPC observed during the resting-state baseline. Visual feedback of γSPC was updated every 40 ms, i.e., at a frame rate of 25 Hz. Online signal processing and visual feedback was implemented in BCI2000 [43] and its extension BCPy2000 (http://bci2000. org/downloads/BCPy2000/).

and then separated it into independent components (ICs) using the SOBI-algorithm [47]. We then inspected the topography, spectrum, and time-series of each IC, and rejected those ICs as non-cortical for which at least one of the following four criteria applied [36]: (1) The spectrum did not show the 1 f -behaviour typical of a cortical source. In particular, we rejected ICs that displayed a monotonic increase in spectral power starting around 20 Hz, as this is characteristic of muscular activity [33]. (2) Eye blinks were detectable in the time-series. (3) The topography did not show a dipolar pattern. (4) The time-series appeared to be contaminated by other noise sources such as 50 Hz line noise or large spikes. We then reprojected the remaining ICs to the scalp. While this procedure attenuates the extent of EMG confounding, it would be unreasonable to expect a full elimination of EMG activity. As described in detail in the following sub-section, we thus compared the decoding accuracy for each subject prior to and following artifact attenuation. We excluded subjects from further analysis if the ICA-based attenuation of EMG activity resulted in a decrease in decoding accuracy. Conversely, we argue that if an attenuation of EMG signals does not impede—or even enhance—decoding accuracy, it is unlikely that this subject employed muscular activity to control the online feedback. We describe how we validated this artifact-attenuation procedure in section 2.4.3.

2.4. Offline data analysis

Following the online training sessions, we analyzed the recorded data offline to identify and attenuate potential confounding by EMG activity (section 2.4.1), analyzed the discriminability of up- and down-regulation of γSPC (section 2.4.2), and investigated the spectral and spatial specificity of the observed effects (section 2.4.3).

2.4.2. Decoding. We used linear discriminant analysis to

EEG recordings in the γrange are likely to be contaminated by EMG activity arising from scalp muscles [33, 44]. While source localization methods such as beamforming are robust against a moderate contamination by EMG activity [37], subjects might have been able to influence the feedback signal by altering the tonus of scalp muscles. In order to identify and attenuate potential confounding by EMG activity, we employed independent component analysis (ICA) [45, 46]: we first high-pass filtered each subjectʼs data using a third order Butterworth filter with a cut-off frequency of 3 Hz, reduced it to 64 dimensions by principal component analysis,

estimate the accuracy in discriminating up- and downregulation of γSPC , with trial-averaged γSPC as the onedimensional feature space. In order to avoid overfitting effects, we employed a leave-one-trial-out cross-validation procedure. As subjects had to explore how to self-regulate γSPC during the first 20-min feedback block, only trials from the last two blocks were used for classification. The classification procedure was carried out on each subjectʼs data once before and once following ICA-based artifact attenuation. Subjects were only kept for further analysis if the artifact attenuation procedure did not result in a decrease in cross-validated decoding accuracy.

2.4.1. Attenuation of EMG artifacts.

4

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

control, allowing her to answer yes/no questions by moving her eyes to the left or to the right. Communication by gaze control was used to obtain informed consent, in agreement with guidelines of the Max Planck Society. Due to the low level of gaze control, answering a question required roughly ten seconds. Oculomotor control rapidly diminished from question to question, requiring her to rest after about five questions before attempting further answers. LEK participated in 18 feedback sessions, carried out at the same time of day and on the same day of the week. While the first four feedback sessions were carried out in consecutive weeks, there was a disease-related intermission of four months before the last 14 sessions could be recorded. LEK quit five sessions prematurely due to mucus building up within her throat, causing painful pharyngeal reflexes. These sessions were excluded from the analysis. All recordings were carried out in the patientʼs home.

2.4.3. Spectral and spatial specificity of bandpower modulation. To investigate the spectral specificity of

bandpower modulation in the SPC, we computed r2-values for each subjectʼs beamforming signal across the whole frequency range, using a sliding window of 3 Hz width with center-frequencies ranging from two to 249 Hz. These r2values range from zero to one and quantify the amount of variance in each frequency band across trials that can be explained by the class labels (up- or down-regulation of γSPC ). We then multiplied each r2-value by the sign of the correlation between class-labels and trial-averaged bandpower, such that positive r2-values indicate an increase in bandpower when subjects are asked to up-regulate γSPC and a decrease in bandpower when subjects are asked to downregulate it. We refer to this measure as the signed r2. To investigate whether subjects learned to focus selfregulation of γ-power to the SPC, we used a beamformer to scan each subjectʼs brain for γ-modulation. Specifically, we successively pointed a LCMV-beamformer at each of the K cortical dipoles modeled by the leadfield matrix L (see section 2.3) and estimated the signed r2-value of the resulting signal in the γ-range (55–85 Hz). We thereby obtained a cortical map that displays, for each subject, cortical locations at which γ-power was modulated in accordance with the experimental conditions. We refer to this map as the cortical r2-map. In order to validate the ICA-based artifact attenuation procedure we repeated this scanning procedure. Instead of retaining the non-artifactual ICA-components, however, we only reprojected those ICA-components that we identified as being of EMG origin (see section 2.4.1). In this way, we obtained a cortical map of signed r2-values that indicates how our estimates of γ-modulation are confounded by EMG activity. We refer to this map as the EMG r2-map. We argue that observed effects should only be interpreted as being of cortical origin if they appear in the cortical but not in the EMG r2-map. As we argued in section 1, modulation of γSPC is likely to be linked to attentional processes. As such, it is reasonable to expect that up- and down-regulation of γSPC also affects various other neural processes. In particular, we expected selfregulation of γSPC to modulate SMRs [36]. To investigate this issue we scanned the spatial distribution of signed r2-values, in the same manner as described earlier in this sub-section, in the θ- (4–8 Hz), α μ- (8–12 Hz), and β-range (20–30 Hz).

2.5.2. Online feedback.

As LEK retained very little oculomotor control, the visual feedback was replaced by feedback in the auditory domain. Two publicly available sound samples were selected from the freesound repository for this purpose (http://freeesound.org). The first sample consisted of a group of people humming (http://freesound. org/people/freesound/sounds/50168/). The second sample was recorded in a forest with wind rustling in the trees and birds chirping (http://freesound.org/people/homejrande/ sounds/17383/). Both samples were edited to play as continuous loops at constant volume. During the feedback sessions, the current level of γSPC was mapped to the volume of these sound samples in analogy to the ballʼs position used for visual feedback (see section 2.1). The humming sound was only played when γSPC exceeded its mean during the resting-base baseline, with the volume linearly increasing from zero to its maximum value when γSPC reached two positive standard deviations of resting-state γSPC. Conversely, the wind-rustling sound was only played when γSPC fell below its mean during the resting-base baseline, with the volume linearly increasing from zero to its maximum value when γSPC reached two negative standard deviations of resting-state γSPC . Instructions to the subject were recorded by a male speaker and automatically read out by the computer system when appropriate. The mental strategies employed by the healthy subjects to control γSPC (see section 3.4) were read out to LEK prior to the first feedback session. Following the fourth feedback session, the length of each trial was reduced from 60 s to 30 s, while the number of trials per block was increased from 20 to 30. This resulted in an increase in the total number of trials per session from 60 to 90, while slightly reducing the time needed to complete each session.

2.5. Training of a locked-in ALS patient 2.5.1. Subject. Following the training sessions with the

healthy subjects, we recruited one female locked-in ALS subject for this study. At the time of the recordings this patient, subsequently referred to as LEK, was 53 years old. She was diagnosed with ALS 15 years ago. On the revised ALS functional rating scale (ALS-FRS-R), a metric to quantify disease progression [48], she scored zero out of a maximum of 48 points. She had no control of her limbs, was artificially ventilated, and was provided with nutrition via a PEG tube. LEK did, however, retain intermittent oculomotor

3. Experimental results We first describe the experimental results obtained with the healthy subjects in sections 3.1 to 3.5. The decoding results achieved with the locked-in ALS patient are reported in section 3.6. Statistical tests are only reported for the initial 5

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

3.2. Spectral and spatial specificity of bandpower modulation

Figure 4 displays the group-average spectral r2-profile of bandpower modulation in the SPC (see section 2.4.3). This plot indicates that subjects employed a broad-band modulation in the γ-range to control the feedback signal with a peak r2-value at 121 Hz. Figure 5 displays group-average cortical maps of r2values in the γ-band (see section 2.4.3). A distinct modulation of γ-power is evident in a fronto-parietal network with a focus in the SPC, which is in agreement with the region targeted by the beamformer (see figure 3). Importantly, this network of γmodulation is only evident in the cortical- but not in the EMG r2-map (figure 6), providing empirical support for a cortical origin of γ-modulation in the SPC. This is in contrast to modulation of γ-power in the occipital cortex, which is found in both the cortical- and the EMG r2-map. We suspect this to result from a low coverage of the occipital cortex by EEG

Figure 4. Group-average spectral profile of r2-values in the superior

parietal cortex, displaying a bandpower modulation in a broad γrange (see section 2.4.3 for details).

Table 1. Decoding performance of healthy subjects following a 20 min training session in percentage of correctly decoded trials.

Subject Accuracy (%)

S1

S2

S3

S4

S5

S6

S7

S8

S9

S10

S11

47.5

72.5

95.0

60.0

92.5

92.5

55.5

80.0

62.5

57.5

57.5

hypothesis that subjects can learn to regulate γSPC . Results of exploratory post-hoc analyses are discussed qualitatively.

channels at the edge of the EEG cap, resulting in a poor attenuation of EMG artifacts generated by neck muscles.

3.1. Decoding results

3.3. Secondary effects of γ SPC -modulation in other frequency bands

Consistent with our previous findings on the relationship of γSPC and SMRs [36], we found that instructing subjects to upand down-regulate γSPC led to a bilateral decrease and increase of μ-rhythms in sensorimotor areas, respectively (figure 7). In addition, the amplitude of frontal α-rhythms decreased when subjects were instructed to up-regulate γSPC and increased when they were told to down-regulate it. Both secondary effects are not evident in the corresponding EMG r2-map (figure 8), indicating that estimates of bandpower modulation in the α μ-range are not confounded by EMG activity. We found no consistent effects of γSPC-regulation across subjects in the θ- or the β-band.

The group-average gain vector g of the beamformer is shown in figure 3 (see section 2.3). This figure indicates that, as intended, the beamforming procedure focused on γ-power originating in the right SPC. For eight of the 19 subjects decoding accuracy decreased due to the artifact-attenuation procedure (see sections 2.4.1 and 2.4.2). These subjects were excluded from further analyses due to potential contamination of the feedback signal by EMG activity. Decoding results obtained for the remaining eleven subjects are shown in the second column of table 1. On a group-level, subjects scored an average classification accuracy of 70.2%. Single-subject results ranged from 47.5% to 95%. A Wilcoxon signed rank test rejected the nullhypothesis of median classification accuracy of 50% (chancelevel) across subjects at p = 0.002. On the individual subject level, a decoding accuracy of 62.5% (70% with Bonferroni correction for multiple (11) comparisons) would be required for the present setup to reject the null-hypothesis of chancelevel accuracy at a significance-level of α = 0.05 [49]. However, above chance-level decoding accuracy is of little use in brain–computer interfacing, as a classification accuracy >70% is usually considered necessary for communication. Five out of eleven subjects achieved this level of performance after a training period of 20 min.

3.4. Subject reports

Following the feedback training, subjects were given a questionnaire in which they were asked to describe how they controlled the feedback signal and note down any feelings induced by up- and down-regulation of γSPC. As neurofeedback is an implicit form of learning, these reports should be interpreted with caution. Subjects explored a wide range of strategies, ranging from various visual imagery tasks to alternating between different attentional states. Only reports of subjects with a decoding accuracy greater than 70% are discussed here (S2, S3, S5, S6 and S8). Subject S2 noted that she simply moved the feedback ball into the desired direction, as is common in 6

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

Figure 5. Group-average cortical r2-map in the γ-range (55–85 Hz) (see section 2.4.3 for details). Modulation of cortical γ-power is strongest

in the superior parietal cortex and extends to more frontal regions.

Figure 6. Group-average electromyogenic r2-map in the γ-range (55–85 Hz) (see section 2.4.3 for details). Contamination of γ-range modulation in the visual cortex by EMG activity is likely a result of sparse coverage of occipital areas by EEG electrodes.

2

Figure 7. Group-average cortical r -map in the α μ -range (8–12 Hz) (see section 2.4.3 for details). Self-regulation of γSPC results in a bilateral

modulation of sensorimotor rhythms. Note the opposite sign of the color scheme relative to figures 5 and 6.

neurofeedback paradigms [50]. Subjects S3, S6 and S8, however, stated that they alternated between states of focused attention, e.g. involving a complex ice-skating choreography (S8), and relaxed wakefulness, e.g. by trying not to think (S3 and S6), to enhance and decrease γSPC, respectively. Subject S5 indicated that she imagined driving up and down within my body to increase or decrease γSPC.

experimental setup, all three blocks of each session (60 trials) were used for decoding. Subject S3 consistently performed at or above 90% accuracy in all feedback sessions. This high level of performance is also reflected in the r2-profile of bandpower modulation in the SPC (figure 9, left column), showing a consistent modulation in the 55–85 Hz feedback range. In terms of the cortical r2-map, subject S3 generated a broad fronto-parietal modulation of bandpower in the γ-range during the last two feedback sessions (figure 10). In all five feedback sessions, subject S3 indicated that she alternated between

3.5. Results of additional training sessions

Results of decoding for the additional four training sessions are listed in table 2. As subjects were already familiar with the 7

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

2

Figure 8. Group-average electromyogenic r -map in the α μ -range (8–12 Hz), indicating that modulation of sensorimotor rhythms shown in

figure 7 is not confounded by EMG activity (see section 2.4.3 for details).

3.6. Decoding results of the locked-in ALS patient

Table 2. Classification performances in percent of correct trials for the additional training sessions.

Subject S3 S8 S10

Session 2

Session 3

Session 4

Session 5

90.0 61.7 33.3

95.0 63.3 46.7

95.0 45.0 56.7

91.7 43.3 61.7

In five out of the 13 training sessions completed by LEK, decoding accuracy decreased due to the ICA-based artifactattenuation procedure (see sections 2.4.1 and 2.4.2). These sessions were excluded from further analysis due to potential confounding by EMG activity. Offline decoding results for the remaining eight sessions are listed in table 3. LEK achieved an average decoding accuracy of 55.8%. A Wilcoxon signed rank test rejected the null-hypothesis of median classification accuracy of 50% (chance-level) across sessions at p = 0.012. LEK did not, however, reach a decoding accuracy greater than 70%, which is considered necessary for communication, in any of the training sessions. Before discussing the spectral and spatial r2-profiles for LEK, we would like to point out that these measures, contrary to the beamforming signal used for decoding, do not integrate over a wide frequency range and a large cortical area. In combination with the low decoding accuracy achieved by LEK, we found this to result in r2-profiles with a very low signal-to-noise ratio (SNR). We have thus chosen to present LEKʼs r2-profiles averaged over the three best-performing sessions only (sessions 1, 2 and 10). The resulting r2-profiles should be interpreted with the appropriate caution. Figure 13 displays LEKʼs spectral r2-profile in the SPC. This profile exhibits a similar shape to the ones found in healthy subjects after repeated training sessions, with a positive peak in the γ- and a negative peak in the α-range (see figure 9). In terms of the cortical r2-map in the γ-range, we again found a fronto-parietal modulation of γ-power (figure 14). In contrast to the healthy subjects, however, LEK displayed a focus of γ-modulation in frontal rather than in parietal areas. This modulation pattern is apparent only in the cortical- and not in the EMG r2-map (figure 15). Despite the low SNR, this supports the interpretation that in the best three sessions LEK was capable of self-regulating fronto-parietal γpower. Consistent with our initial argument that SMRs are impaired in ALS, we found no spatially coherent effects in LEKʼs cortical r2-maps in the μ-range (not shown).

states of relaxed wakefulness and focused attention to regulate γSPC down and up, respectively. Subject S8 performed reasonably well in the first three feedback sessions (see tables 1 and 2), alternating between relaxed wakefulness and focusing on complex tasks to decrease and increase γSPC , respectively. During subsequent training sessions, however, S8 indicated that she changed her cognitive control strategy to motor-imagery (having previously participated in a motor-imagery BCI study). This was accompanied by a drop in decoding accuracy to chance-level table (2) and the absence of distinct γ-modulation in the SPC (figure 9, middle column) or any other part of cortex (figure 11). Subject S10 initially proved to be incapable of self-regulating γSPC . With repeated training, however, S10 learned to control γSPC, resulting in a decoding accuracy of 61.7% in the last training session. This learning effect is also reflected in the r2-profile of bandpower modulation in the SPC (figure 9, right column), leading to a fronto-parietal γ-modulation in the last two training sessions (figure 12) that is consistent with the ones observed on the group-level and in subject S3 (see figures 5 and 10). In terms of cognitive control strategies, S3 indicated after every training session that she did not use any explicit control strategy. Over the course of the five feedback sessions, all three subjects developed secondary effects of γSPC -regulation in various other frequency bands, including the θ-, α μ- and βband. We did not find these effects to be consistent across subjects, however, potentially due to different cognitive control strategies. An in-depth analysis of the subject-specific effects of γSPC -regulation on other frequency bands is beyond the scope of the present article. 8

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

Figure 9. Spectral profiles of r2-values in the superior parietal cortex for all training sessions of subjects S3, S8 and S10 (see section 2.4.3 for

details).

Figure 10. Cortical r2-map of subject S3, averaged over the last two (fourth and fifth) training sessions (see section 2.4.3 for details).

4. Discussion

subjectʼs (S8) change in cognitive control strategies from attention-regulation to motor-imagery resulted in a loss of the ability to self-regulate γSPC. Our interpretation, that γSPC in an indicator of a high-level cognitive process, is further supported by previous reports on the relevance of the SPC [54] as well as of γ-oscillations [55, 56] for top–down attentional control. While several BCI systems based on attentional paradigms have been proposed in the past, these paradigms require subjects to modulate the brainʼs response to external stimuli [16, 20–22, 57]. Even though the paradigm presented here also relies on sensory feedback during the initial training phase, the actual transmission of information from the subject to the BCI does not require sensory processing. Regarding the performance of our system, we note that while the group-average decoding accuracy of 70.2% is comparable to a simple motor-imagery BCI [58], the long trial-time of our paradigm results in a low information transfer rate. As we have indicated in the introduction, however, we did not aim to design a BCI with a high information transfer rate. Instead, our goal was to establish a novel paradigm that will allow us to investigate whether CLI patients in late stages of ALS retain the capacity for goal-directed thinking. As such, the crucial question is whether subjects in late stages of ALS are able to modulate γSPC . While only a small percentage of healthy subjects appear incapable of operating a BCI [58, 59], successful decoding results with severely paralyzed ALS patients are scarce. Only two ALS patients with a severity of impairment comparable to that of LEK have been

We argued in the introduction that in order to challenge the thought-extinction hypothesis in ALS patients, it is essential to develop novel BCI paradigms that directly target high-level cognitive processes. In the following, we discuss the available evidence that self-regulation of γSPC is indeed associated with high-level cognitive control and consider the significance of our findings. In previous work, we have shown that γ-power during rest in a fronto-parietal network including the SPC predicts the ability of subjects to modulate their sensorimotor μrhythm [36]. This led us to hypothesize that γ-power in the SPC is an indicator of information processing in a frontoparietal attention network that modulates SMRs [38]. In agreement with this hypothesis, we found here that self-regulation of γSPC leads to a bilateral modulation of μ-rhythms in the sensorimotor cortex (see figure 7). Importantly, we did not find any evidence for a distinct modulation of γ-power in sensorimotor areas, which is known to accompany changes in μ-power when subjects engage in motor tasks or motor-imagery [34, 51–53] (see figure 5). We interpret this as evidence that the observed changes in μ-power are not movementrelated but are a result of self-regulation of γSPC . This interpretation is consistent with the subjects’ reports on their cognitive control strategies. Three out of the five most successful subjects alternated between states of focused attention and relaxed wakefulness to control γSPC. Furthermore, one 9

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

Figure 11. Cortical r2-map of subject S8, averaged over the last two (fourth and fifth) training sessions (see section 2.4.3 for details).

Figure 12. Cortical r2-map of subject S10, averaged over the last two (fourth and fifth) training sessions (see section 2.4.3 for details).

that throughout all training sessions residual eye-movement was the most reliable form of communication. It thus remains an open question whether ALS patients in general and LEK in particular may benefit from our work when entering the CLI state. This raises the question of how our system may be further improved. Firstly, we note that computation of γSPC was based on a group-level template derived from previous work [36]. Substantially higher performances may be expected by learning subject-specific spatial filters that also incorporate frequencies beyond the γ-range [61, 62]. Secondly, the long trial-length limits the number of training trials that can be recorded in one session, thereby restricting the maximum size of the feature space. Addressing this issue may require more advanced machine-learning techniques such as multi-task learning to exploit structure in the feature space that is shared across sessions [63]. Thirdly, the motivational aspect of operant conditioning has largely been neglected in this study [64, 65]. A feedback scheme that provides more engaging rewards may result in enhanced training effects. Finally, we note that LEK was already in the locked-in state at the start of training. In order to improve the chances of falsifying the thought-extinction hypothesis, we consider it essential to start training patients at an early stage and accompany them throughout disease progression into the CLI state.

2

Figure 13. Spectral profile of r -values in the superior parietal cortex

of ALS patient LEK, averaged across the first, second, and tenth training session (see sections 2.4.3 and 3.6 for details). While bandpower modulation is very weak in LEK, the spectral profile resembles that of healthy subjects with a positive peak in the feedback range (55–85 Hz) and negative values in the α-range.

reported to achieve above chance-level decoding accuracy in a BCI task [8]. In a recent study with three locked-in ALS patients, none learned to successfully operate the BCI [60]. This underscores the importance of developing novel BCI paradigms for patients in late stages of ALS. While the decoding results obtained by LEK are encouraging, we note 10

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

Table 3. Decoding results of eight recording sessions with the locked-in ALS patient (see section 3.6). Accuracy is given in percentage of

correctly decoded trials. Session Accuracy Trials

1

2

3

6

10

13

15

18

56.7% 60

60.0% 60

55.0% 60

55.6% 90

57.8% 90

55.6% 90

50.0% 90

55.6% 90

Average 55.8% 630

Figure 14. Cortical r2-map of LEK in the γ-range, averaged over the first, second, and tenth training session (see sections 2.4.3 and 3.6 for

details). Despite very low r2-values, a coherent pattern of γ-modulation is apparent in fronto-parietal areas, overlapping with regions also engaged by healthy subjects (see figure 5).

Figure 15. Electromyogenic r -map of LEK in the γ-range, averaged over the first, second, and tenth training session (see sections 2.4.3 and 2

3.6 for details). In contrast to the cortical r2-map in the γ-range (see figure 14), no coherent pattern of γ-modulation is apparent.

References

[7] Birbaumer N, Piccione F, Silvoni S and Wildgruber M 2012 Ideomotor silence: the case of complete paralysis and braincomputer interfaces (BCI) Psychological Res. 76 183–91 [8] Kübler A and Birbaumer N 2008 Brain–computer interfaces and communication in paralysis: extinction of goal directed thinking in completely paralysed patients? Clin. Neurophysiol. 119 2658–66 [9] Mason S G, Bashashati A, Fatourechi M, Navarro K F and Birch G E 2007 A comprehensive survey of brain interface technology designs Ann. Biomed. Eng. 35 137–69 [10] Nihei K, McKee A C and Kowall N W 1993 Patterns of neuronal degeneration in the motor cortex of amyotrophic lateral sclerosis patients Acta Neuropathologica 86 55–64 [11] Mai R, Facchetti D, Micheli A and Poloni M 1998 Quantitative electroencephalography in amyotrophic lateral sclerosis Electroencephalogr. Clin. Neurophysiol. 106 383–6 [12] Nijboer F et al 2008 A P300-based brain–computer interface for people with amyotrophic lateral sclerosis Clin. Neurophysiol. 119 1909–16

[1] Vidal J J 1973 Toward direct brain–computer communication Annu. Rev. Biophys. Bioeng. 2 157–80 [2] Wolpaw J R, McFarland D J, Neat G W and Forneris C A 1991 An EEG-based brain–computer interface for cursor control Electroencephalogr. Clin. Neurophysiol. 78 252–9 [3] Birbaumer N, Ghanayim N, Hinterberger T, Iversen I, Kotchoubey B, Kübler A, Perelmouter J, Taub E and Flor H 1999 A spelling device for the paralysed Nature 398 297–8 [4] Donchin E, Spencer K M and Wijesinghe R 2000 The mental prosthesis: assessing the speed of a P300-based brain–computer interface IEEE Trans. Rehabil. Eng. 8 174–9 [5] Pfurtscheller G and Neuper C 2001 Motor imagery and direct brain-computer communication Proc. IEEE 89 1123–34 [6] Krusienski D J, Grosse-Wentrup M, Galán F, Coyle D, Miller K J, Forney E and Anderson C W 2011 Critical issues in state-of-the-art brain–computer interface signal processing J. Neural Eng. 8 025002 11

J. Neural Eng. 11 (2014) 056015

M Grosse-Wentrup and B Schölkopf

[35] Darvas F, Scherer R, Ojemann J G, Rao R P, Miller K J and Sorensen L B 2010 High gamma mapping using EEG NeuroImage 49 930–8 [36] Grosse-Wentrup M and Schölkopf B 2012 High gamma-power predicts performance in sensorimotor-rhythm brain–computer interfaces J. Neural Eng. 9 046001 [37] Hipp J F and Siegel M 2013 Dissociating neuronal gammaband activity from cranial and ocular muscle activity in EEG Front. Hum. Neurosci. 7 338 [38] Grosse-Wentrup M and Schölkopf B 2013 A review of performance variations in SMR-based brain–computer interfaces (BCIs) Brain–Computer Interface Research ed C Guger, B Z Allison and G Edlinger (Berlin: Springer) pp 39–51 [39] Maquet P et al 2005 Human cognition during REM sleep and the activity profile within frontal and parietal cortices: a reappraisal of functional neuroimaging data Prog. Brain Res. 150 219–72 [40] van Veen B D, van Drongelen W, Yuchtman M and Suzuki A 1997 Localization of brain electrical activity via linearly constrained minimum variance spatial filtering IEEE Trans. Biomed. Eng. 44 867–80 [41] Grosse-Wentrup M, Liefhold C, Gramann K and Buss M 2009 Beamforming in non-invasive brain–computer interfaces IEEE Trans. Biomed. Eng. 56 1209–19 [42] Mosher J C, Baillet S, Darvas F, Pantazis D, Yildirim E K and Leahy R M 2005 Brainstorm electromagnetic imaging software 5th Int. Symp. on Noninvasive Functional Source Imaging within the Human Brain and Heart (NFSI 2005) [43] Schalk G, McFarland D J, Hinterberger T, Birbaumer N and Wolpaw J R 2004 BCI2000: a general-purpose brain–computer interface (BCI) system IEEE Trans. Biomed. Eng. 51 1034–43 [44] Shackman A J, McMenamin B W, Slagter H A, Maxwell J S, Greischar L L and Davidson R J 2009 Electromyogenic artifacts and electroencephalographic inferences Brain Topography 22 7–12 [45] Comon P 1994 Independent component analysis, a new concept? Signal Process. 36 287–314 [46] McMenamin B W, Shackman A J, Maxwell J S, Bachhuber D R W, Koppenhaver A M, Greischar L L and Davidson R J 2010 Validation of ICA-based myogenic artifact correction for scalp and source-localized EEG NeuroImage 49 2416–32 [47] Belouchrani A, Abed-Meraim K, Cardoso J-F and Moulines E 1997 A blind source separation technique using secondorder statistics IEEE Trans. Signal Process. 45 434–44 [48] Cedarbaum J M, Stambler N, Malta E, Fuller C, Hilt D, Thurmond B and Nakanishi A 1999 The ALSFRS-R: a revised ALS functional rating scale that incorporates assessments of respiratory function J. Neurological Sci. 169 13–21 [49] Müller-Putz G R, Scherer R, Brunner C, Leeb R and Pfurtscheller G 2008 Better than random? A closer look on BCI results Int. J. Bioelectromagnetism 10 52–55 [50] Kober S E, Witte M, Ninaus M, Neuper C and Wood G 2013 Learning to modulate oneʼs own brain activity: the effect of spontaneous mental strategies Front. Hum. Neurosci. 7 695 [51] Schalk G, Miller K J, Anderson N R, Wilson J A, Smyth M D, Ojemann J G, Moran D W, Wolpaw J R and Leuthardt E C 2008 Two-dimensional movement control using electrocorticographic signals in humans J. Neural Eng. 5 75 [52] Pistohl T, Ball T, Schulze-Bonhage A, Aertsen A and Mehring C 2008 Prediction of arm movement trajectories from ECoG-recordings in humans J. Neurosci. Methods 167 105–14 [53] Gomez-Rodriguez M, Grosse-Wentrup M, Peters J, Naros G, Hill J, Schölkopf B and Gharabaghi A 2010 Epidural ECoG online decoding of arm movement intention in hemiparesis

[13] Brunner P, Joshi S, Briskin S, Wolpaw J R, Bischof H and Schalk G 2010 Does the P300 speller depend on eye gaze? J. Neural Eng. 7 056013 [14] Jacobs L, Bozian D, Heffner R R and Barron S A 1981 An eye movement disorder in amyotrophic lateral sclerosis Neurology 31 1282–1282 [15] van Gerven M, Bahramisharif A, Heskes T and Jensen O 2009 Selecting features for BCI control based on a covert spatial attention paradigm Neural Netw. 22 1271–7 [16] Treder M S, Schmidt N M and Blankertz B 2011 Gazeindependent brain–computer interfaces based on covert attention and feature attention J. Neural Eng. 8 066003 [17] Brouwer A-M and van Erp J B F 2010 A tactile P300 brain–computer interface Front. Neurosci. 4 19 [18] Ramos Murguialday A, Hill N J, Bensch M, Martens S, Halder S, Nijboer F, Schölkopf B, Birbaumer N and Gharabaghi A 2011 Transition from the locked in to the completely locked-in state: a physiological analysis Clin. Neurophysiol. 122 925–33 [19] Nijboer F, Furdea A, Gunst I, Mellinger J, McFarland D J, Birbaumer N and Kübler A 2008 An auditory brain–computer interface (BCI) J. Neurosci. Methods 167 43–50 [20] Schreuder M, Blankertz B and Tangermann M 2010 A new auditory multi-class brain–computer interface paradigm: spatial hearing as an informative cue PLoS One 5 e9813 [21] Hill N J and Schölkopf B 2012 An online brain–computer interface based on shifting attention to concurrent streams of auditory stimuli J. Neural Eng. 9 026011 [22] Kübler A, Furdea A, Halder S, Hammer E M, Nijboer F and Kotchoubey B 2009 A brain–computer interface controlled auditory event-related potential (P300) spelling system for locked-in patients Ann. New York Acad. Sci. 1157 90–100 [23] Hill N J, Ricci E, Haider S, McCane L M, Heckman S, Wolpaw J R and Vaughan T M 2014 A practical, intuitive brain–computer interface for communicating ‘yes’ or ‘no’ by listening J. Neural Eng. 11 035003 [24] Kübler A, Kotchoubey B, Salzmann H-P, Ghanayim N, Perelmouter J, Hömberg V and Birbaumer N 1998 Selfregulation of slow cortical potentials in completely paralyzed human patients Neurosci. Lett. 252 171–4 [25] Birbaumer N 1999 Slow cortical potentials: plasticity, operant control, and behavioral effects The Neuroscientist 5 74–78 [26] He B J and Raichle M E 2009 The fMRI signal, slow cortical potential and consciousness Trends Cogn. Sci. 13 302–9 [27] Koch C 2009 The SCP is not specific enough to represent conscious content Proc. Natl Acad. Sci. USA 100 8520–5 [28] Wang X-J and Buzsáki G 1996 Gamma-oscillation by synaptic inhibition in a hippocampal interneuronal network model J. Neurosci. 16 6402–13 [29] Buzsaki G 2006 Rhythms of the Brain (Oxford: Oxford University Press) [30] Bartos M, Vida I and Jonas P 2007 Synaptic mechanisms of synchronized gamma-oscillations in inhibitory interneuron networks Nat. Rev. Neurosci. 8 45–56 [31] Jensen O, Kaiser J and Lachaux J P 2007 Human gammafrequency oscillations associated with attention and memory Trends Neurosci. 30 317–24 [32] Fries P, Nikolić D and Singer W 2007 The gamma cycle Trends Neurosci. 30 309–16 [33] Goncharova I I, McFarland D J, Vaughan T M and Wolpaw J R 2003 EMG contamination of EEG: Spectral and topographical characteristics Clin. Neurophysiol. 14 1580–93 [34] Ball T, Demandt E, Mutschler I, Neitzel E, Mehring C, Vogt K, Aertsen A and Schulze-Bonhage A 2008 Movement related activity in the high gamma range of the human EEG NeuroImage 41 302–10

12

J. Neural Eng. 11 (2014) 056015

[54] [55]

[56]

[57]

[58]

[59]

M Grosse-Wentrup and B Schölkopf

First Workshop on Brain Decoding: Pattern Recognition Challenges in Neuroimaging (WBD) (Piscataway, NJ: IEEE) pp 36–39 Behrmann M, Geng J J and Shomstein S 2004 Parietal cortex and attention Curr. Opin. Neurobiol. 14 212–7 Ray S, Niebur E, Hsiao S S, Sinai A and Crone N E 2008 High-frequency gamma activity (80–150Hz) is increased in human cortex during selective attention Clin. Neurophysiol. 119 116–33 Gunduz A, Brunner P, Daitch A, Leuthardt E C, Ritaccio A L, Pesaran B and Schalk G 2012 Decoding covert spatial attention using electrocorticographic (ECoG) signals in humans NeuroImage 60 2285–93 Allison B Z, McFarland D J, Schalk G, Zheng S D, Jackson M M and Wolpaw J R 2008 Towards an independent brain–computer interface using steady state visual evoked potentials Clin. Neurophysiol. 119 399–408 Guger C, Daban S, Sellers E, Holzner C, Krausz G, Carabalona R, Gramatica F and Edlinger G 2009 How many people are able to control a P300-based brain–computer interface (BCI)? Neurosci. Lett. 462 94–98 Guger C, Edlinger G, Harkam W, Niedermayer I and Pfurtscheller G 2003 How many people are able to operate

[60] [61]

[62] [63] [64] [65]

13

an EEG-based brain–computer interface (BCI)? IEEE Trans. Neural Syst. Rehabil. Eng. 11 145–7 de Massari D, Ruf C A, Furdea A, Matuz T, van der Heiden L, Halder S, Silvoni S and Birbaumer N 2013 Brain communication in the locked-in state Brain 136 1989–2000 Blankertz B, Dornhege G, Krauledat M, Müller K R and Curio G 2007 The non-invasive berlin brain–computer interface: fast acquisition of effective performance in untrained subjects NeuroImage 27 539–50 Grosse-Wentrup M and Buss M 2008 Multiclass common spatial patterns and information theoretic feature extraction IEEE Trans. Biomed. Eng. 55 1991–2000 Alamgir M, Grosse-Wentrup M and Altun Y 2010 Multitask learning for brain–computer interfaces Proc. 13th Int. Conf. on Artificial Intelligence and Statistics (AISTATS) pp 17–24 Fetz E E 2007 Volitional control of neural activity: implications for brain–computer interfaces J. Physiol. 579 571–9 Lotte F, Larrue F and Mühl C 2013 Flaws in current human training protocols for spontaneous brain–computer interfaces: lessons learned from instructional design Front. Hum. Neurosci. 7 568