wireless eeg with individualized channel layout enables efficient motor imagery training
TRANSCRIPT
Accepted Manuscript
Wireless EEG with individualized channel layout enables efficient motor im-agery training
Catharina Zich, Maarten De Vos, Cornelia Kranczioch, Stefan Debener
PII: S1388-2457(14)00370-8DOI: http://dx.doi.org/10.1016/j.clinph.2014.07.007Reference: CLINPH 2007167
To appear in: Clinical Neurophysiology
Accepted Date: 7 July 2014
Please cite this article as: Zich, C., Vos, M.D., Kranczioch, C., Debener, S., Wireless EEG with individualizedchannel layout enables efficient motor imagery training, Clinical Neurophysiology (2014), doi: http://dx.doi.org/10.1016/j.clinph.2014.07.007
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customerswe are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, andreview of the resulting proof before it is published in its final form. Please note that during the production processerrors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Wireless EEG with individualized channel layout enables efficient
motor imagery training
Catharina Zich1, Maarten De Vos
2,3,4, Cornelia Kranczioch
1,3, Stefan Debener
1,3,4
1 Neuropsychology Lab, Department of Psychology, European Medical School, Carl von Ossietzky
University of Oldenburg, Germany
2 Methods in Neurocognitive Psychology, Department of Psychology, European Medical School, Carl
von Ossietzky University of Oldenburg, Germany
3 Neurosensory Science Research Group, Carl von Ossietzky University of Oldenburg, Germany
4 Cluster of Excellence Hearing4all, Carl von Ossietzky University of Oldenburg, Germany
Corresponding Author:
Catharina Zich
Department of Psychology
University of Oldenburg
26111 Oldenburg, Germany
Tel: +49 441 798 2172
Fax: +49 441 798 5522
E-mail: [email protected]
Keywords: Motor imagery, Mobile EEG, Brain-computer interface, Electrode reduction.
Highlights
1. To reduce the number of channels to two bipolar channels a novel reduction method for
motor imagery is presented, which outperformed a published procedure and the standard
placement.
2. Using an individualized electrode subset, a longitudinal study was performed over four
days of motor imagery practice, in which significant learning effects could be observed.
3. On days 2 to 4 practice took place in an everyday environment using a very user-friendly
EEG system consisting of individualized caps and a low-cost wireless hardware.
Abstract
Objective: The study compared two channel-reduction approaches in order to investigate the
effects of systematic motor imagery (MI) neurofeedback practice in an everyday environment
using a very user-friendly EEG system consisting of individualized caps and highly portable
hardware.
Methods: Sixteen BCI novices were trained over four consecutive days to imagine left and
right hand movements while receiving feedback. The most informative bipolar channels for
use on the subsequent days were identified on the first day for each individual based on a
high-density online MI recording.
Results: Online classification accuracy on the first day was 85.1% on average (range: 64.7 -
97.7 %). Offline an individually-selected bipolar channel pair based on common spatial
patterns significantly outperformed a pair informed by independent component analysis and a
standard 10-20 pair. From day 2 to day 4 online MI accuracy increased significantly (day 2:
69.1%; day 4: 73.3%), which was mostly caused by a reduction in ipsilateral event-related
desynchronizantion of sensorimotor rhythms.
Conclusion: The present study demonstrates that systematic MI practice in an everyday
environment with a user-friendly EEG system results in MI learning effects.
Significance: These findings help to bridge the gap between elaborate laboratory studies with
healthy participants and efficient home or hospital based MI neurofeedback protocols.
INTRODUCTION
Motor imagery (MI) electroencephalogram (EEG) neurofeedback training is a promising
therapeutic approach for the recovery of lost motor function. Patients can learn to modulate
the amplitude of sensorimotor rhythms (SMRs) as recorded with electroencephalogram (EEG)
and magnetoencephalography (MEG). This may initiate cortical reorganization and thereby
support functional recovery (Buch et al., 2008; Caria et al., 2011). However,
neurorehabilitation training is in general effortful and a large amount of training over a
prolonged period of time seems necessary to facilitate adaptive cortical reorganization
(Langhorne et al, 2009). Since commonly available EEG technology is rather cumbersome
and not well suited for frequent use outside of the laboratory, the clinical utility of MI
neurofeedback, and in particular the potential benefit it may have for motor rehabilitation,
remains poorly understood. By reducing considerably the number of electrodes and
combining an individualized electrode subset with a small wireless EEG system the present
study aimed to contribute to bridging this gap and to prepare the ground for an MI
neurofeedback system that can efficiently be applied at home or in a rehabilitation setting.
The neurophysiological basis of motor imagery brain-computer interfaces (BCIs) are
amplitudes of SMRs spanning the 8 to 30 Hz frequency range (Pfurtscheller and Neuper,
2001). The modulation of SMRs can be induced by voluntary internal drives, such as the
execution or imagination of movements (Jasper and Penfield, 1949; Pfurtscheller and
Aranibar, 1979; Pfurtscheller et al., 1997; Pfurtscheller et al., 2006). SMR event-related
desynchronization (ERD) patterns seem to be very similar for executed and imagined
movements (McFarland et al., 2000), which supports the idea that motor imagery involves
similar neural mechanisms to those operating during real movements (Jeannerod, 1995;
2001). Since many individuals can voluntarily control their SMR amplitude using motor
imagery as a mental strategy, SMR-based neurofeedback and BCI applications are very
promising novel neurorehabilitation tools (Prasad et al., 2010; Cincotti et al., 2012; Ortner et
al., 2012).
To be useful as neurorehabilitation tools, BCI applications should be suitable for independent
domestic application. This requires affordability, short set-up time, reliability, and usability of
the tools (Sellers et al., 2010; Holz et al., 2013). Achieving these requirements involves
minimizing EEG preparation time. To this end, the number of EEG channels could be
minimized as much as possible. To achieve and optimize this for a MI neurofeedback was the
first main objective of the present study. For right- and left-hand MI paradigms the minimum
number of electrodes is three, supporting one bipolar channel per hemisphere with the same
common midline reference, plus one ground electrode. Here a symmetric electrode layout that
includes the standard electrode positions C3 and C4 (e.g. Kaiser et al., 2011) or adjacent
positions that are approximately located over the primary motor cortex (Wolpaw et al., 2002)
and a midline FCz position as reference are common (e.g., Wang et al., 2006; 2007a).
However, due to individual differences in brain structure and function, a “one cap layout fits
all” strategy may be suboptimal. Hence, it has been argued that individually determined
channel configurations are preferable (Schröder et al., 2005; Lou et al., 2008). An open
question however is how to derive the most optimal electrode sites for each individual. One
possibility would be to compare each potential electrode pair, but because processing
demands would be very high approaches that restrict the search space are favorable. Such
approaches are available, yet to the best of our knowledge systematic comparisons or a
validation over several days are lacking. We therefore aimed not only to reduce the number of
electrodes for an MI neurofeedback to an individually selected bipolar channel pair but,
importantly, to compare two selection approaches.
In this comparison we firstly considered the selection approach proposed by Lou et al. (2008)
that is based on independent component analysis (ICA). Secondly, we suggest and consider a
novel channel selection method. This method is based on common spatial patterns (CSP),
which is currently the state-of-the-art spatial filter in MI BCIs (Blankertz et al., 2008a). The
two selection procedures were compared to a standard symmetrical 10-20 channel pair. We
used inherently independent data obtained on different days in order to verify the individually
selected channels while avoiding circularity. Avoiding circularity is a very important aspect in
cognitive neuroscience in general (Kriegeskorte et al., 2009) and this holds also for the
selection of channel subsets for the long-term use of MI BCIs. To achieve independence we
applied the two channel selection procedures to EEG data collected in a high-density 94-
channel laboratory recording session and tested the selection on independent data from three
consecutive days recorded outside the laboratory.
The reduction of the number of channels considerably reduces EEG preparation time from
about half an hour to five minutes and thus helps to meet the requirements for a BCI-based
rehabilitation tool. Another important aspect of such a tool is the recording hardware, which
should be mobile, affordable and easy to use. We therefore not only reduced the number of
channels, but also changed from a lab-mounted EEG system to a small, wireless and easy to
set up EEG system (Debener et al., 2012; De Vos et al., 2013; De Vos et al., 2014). We also
moved out of the laboratory to mimic the environment of the end-user, while participants
practiced the MI task over three consecutive days. This enabled us to pursue the second main
objective of this study, which was to investigate whether MI learning takes place given the
massive reduction of EEG channels, hardware adaptations, and the weakly controlled
environment. Many studies in the MI BCI field are based on small sample sizes, tend to lack
clear protocols, or mix BCI novices and experienced BCI users. For example Neuper et al.
(1999) trained four MI and BCI experienced subjects over 5 to 10 sessions, scheduled one to
three sessions per week, while Guger et al. (2000) examined three BCI experiences subjects
over three days and McFarland et al. (2010) trained four patients, three with and one without
previous BCI experience, over two to three weeks. However in spite of these shortcomings, a
number of studies have demonstrated that practicing MI tasks that involve EEG-feedback
results in significant learning effects (Neuper et al., (1999), Guger et al., (2000) and Ono et al.
(2013); but also see Neuper et al., 2009; Friedrich et al., 2013), and we similarly expected to
see MI learning effects. To demonstrate that MI learning takes place in spite of the number of
channels, hardware and recording environment as used in our study is essential for all further
efforts to implement a BCI-based rehabilitation tool.
To summarize EEG-based MI neurofeedback is a promising therapeutic approach for the
recovery of lost motor function. However, there is a gap in the way in which research on
EEG-based MI neurofeedback is conducted and the requirements for an EEG-based MI
neurofeedback system that can be efficiently applied at home or in a rehabilitation setting. In
order to reduce this gap we firstly compared two electrode selection approaches for finding
optimal bipolar channels for each individual with a standard channel pair. Secondly, using the
selected channels in combination with a low cost, small and mobile EEG system we
investigated the feasibility of MI learning in an everyday environment.
METHODS
Participants
The original sample consisted of 22 individuals (14 females; age range 18-33 years; mean
25.1 years) with no history of neurological or psychiatric disorders. All participants were right
handed according to the Edinburgh handedness inventory (Oldfield, 1971) and initially naïve
to MI neurofeedback and BCI experiments. Written informed consent was obtained from each
participant. The study was approved by the local ethics committee. After the first day six
participants (3 females, 3 males) were excluded from further participation. One did not follow
task instructions and for five participants the bipolar channel selection strategy was not
applicable (see section Excluded participants for details).
Experimental Paradigm
N=16 individuals participated on all four consecutive days. Recording time was kept constant
for each participant over the four days, to avoid performance fluctuations depending on the
time of day. Every day three experimental sessions (8 min each) were held, one training
session and two feedback sessions. Figure 1 illustrates the study design. Between the sessions,
there were short breaks of 3 to 5 minutes. Individual sessions follow the standard Graz-MI
protocol and consisted of 40 trials (20 left and 20 right hand trials) presented in randomized
order (Pfurtscheller and Neuper, 2001). Note that the machine learning perspective would
recommend many trials within a single session, as this yields a more stable spatial filter and
classifier (e.g. Blankertz et al., 2008a). The study was however designed to accommodate the
end-users’ needs, which are for instance sleepier and suffer from fatigue (Billows et al.,
2013). This requires short MI sessions. It has been recommended for motor recovery after
stroke (Langhorne et al. 2009) and confirmed in clinical MI studies (Kübler et al., 2005; Buch
et al., 2008; McFarland et al., 2010), that motor practice should be repetitive and high-
intensity. In the present study this was accounted for by testing over four consecutive days. In
accordance with this Blankertz and colleagues (2008a) could not observe training effects over
a single long practice session, which led them to conclude that one day of MI practice is not
enough to show learning effects.
Each trial started with a fixation cross in the center of the screen. After 3 s the cross was
overlaid for 1.25 s with a red arrow pointing to the right or left, indicating the beginning of
the motor imagery period. Participants were instructed to imagine the kinesthetic experience
of a sequential finger tapping task (index-middle-ring-little-index-middle-ring-little) from the
first-person perspective with either the right or left hand as indicated by the direction of the
arrow (Neuper et al., 2005). After 8 s the trial ended with a black screen. Inter-trial intervals
ranged from 0.5 to 2.5 s. In the feedback sessions a light blue horizontal bar varying in length
was displayed during the motor imagery interval, representing the real-time classifier output.
The feedback signal was updated at a frequency of 16 Hz. Before the first training and before
the first feedback session, a few sample trials were presented to the participants to ensure they
understood the task. Stimulus presentation was controlled with OpenViBE 0.15.0 (Renard et
al., 2010) running on day one on a local computer (screen diagonal: 24 inch; screen resolution
1920 x 1080 pixels; distance: 160 cm) and on days 2 to 4 on a standard notebook (screen
diagonal: 11.6 inch; resolution 1366 x 768 pixels; distance: 70 cm).
To evaluate the general ability of motor imagery before the first and after the last recording
session, the German Kinesthetic and Visual Imagery Questionnaire (KVIQ-G) was completed
(Schuster et al., 2012) twice. This self-assessment questionnaire is a reliable and valid tool to
evaluate initial motor imagery abilities and possible changes in motor imagery capacity with
practice (Malouin et al., 2007b; Randhawa et al., 2010; Schuster et al., 2012).
--- FIGURE 1 HERE ---
EEG recordings
On the first day, EEG data were acquired from 94 equidistant infra-cerebral scalp sites with a
central fronto-polar site as ground and a nose-tip electrode as reference using a BrainAmp
system (BrainProducts GmbH, Gilching, Germany). Data were recorded with a resolution of
0.1µV and a sampling rate of 500 Hz (0.1 to 250 Hz bandpass). In addition, a surface
electromyogram (EMG) was recorded from two electrodes per hand, placed on the Opponens
Pollicis and First Dorsal Interossei. EMG signals were used to confirm the absence of target
related muscle activity and to ensure that participants followed task instructions. No
participant used EMG activity systematically to operate the BCI. Due to limited channel
capacities EMG was recorded from only one hand.
At days 2 to 4 continuous EEG data were collected with a modified Emotiv
(www.emotive.com) system as described previously (Debener et al., 2012). Briefly, the
original hardware (128 Hz sampling rate; 0.16 and 45 Hz bandpass) was relocated into a small
and light box (49x44x25 mm; 48 grams total weight), which was tightly attached to a state-of-
the-art infra-cerebral electrode cap from Easycap (www.easycap.de) at the back of the head
with an elastic band. The electrode caps had exactly the same 94 openings as the caps used for
the laboratory 94-channel recordings. Sintered Ag/AgCl electrodes in plastic adaptors were
used for the recordings. Five standard electrodes were fixed to the caps: a midline occipital
site as ground and a midline frontopolar electrode as reference; close approximations of the
10-20 electrodes C3, C4 and FCz. The positions of a maximum of eight further electrodes
were determined for each subject individually, after common spatial pattern and ICA analysis
of the 94-channel EEG recordings from the first day. For all electrodes the same reference
was used and bipolar derivations were calculated on- and offline. The resulting channels are
further referred to as bipolar channels. In addition to the scalp channels four EMG electrodes
were positioned, two on the left and two on the right hand. As before EMG data were
recorded for only one hand.
On all four days electrode impedances where kept below 20 kOhm before the recording
started using the BrainAmp Software (BrainProducts GmbH, Gilching, Germany). At the first
day data were acquired in a laboratory environment (i.e. a dimly lit and sound shielded room),
at the following three days data were collected in a rather noisy office room. The room was
chosen with the purpose to achieve recording conditions that corresponded closely to an
everyday environment: day light could not be controlled, noise from a frequently used
adjacent entry hall was clearly audible, and no hardware other than a notebook and the
wireless, mobile EEG device was used (see Fig. 1). While, for the laboratory recording
participants were seated in a comfortable arm chair optimized for EEG recordings, for the
office recordings participants were seated on an ordinary chair. All data were recorded with
OpenViBE. In all recording sessions participants’ hands rested on their lap with palms up.
Participants were asked to keep their arms and hands relaxed and to minimize eye movements
during the recordings.
Data analysis
Data were analyzed offline using EEGLAB 10.2.2.4b (Delorme and Makeig, 2004) and
BCILAB v1.0 (Delorme et al., 2011), running under MATLAB R2011b (Mathworks, Inc.,
Natick, MA).
CSP-informed channel selection
Because CSP is computationally efficient and maximizes the variance of one class while
simultaneously minimizes the variance for the second, CSP has been established as the state-
of-the-art spatial filter in MI applications. The outcome of the CSP is a square matrix W (filter
x channels), which contains the unmixing weights of spatial filter coefficients. As the first and
the last rows of the matrix represent the most discriminative filter with respect to variance, we
used those to select the most discriminative bipolar channel pair. From each filter one bipolar
channel was derived, consisting of the channels with the absolute maximum and minimum
weight. We reasoned that given the restriction of the number of channels to two, these two
would represent the filter best. Initially raw data were bandpass filtered between 8-30 Hz
(Müller-Gerking et al., 1999; Blankertz et al., 2008a) and segmented from -1.0 s before to 4.0
s after onset of the red arrow. Segments containing large artifacts were rejected (on average
13/120 epochs). The interval from 0.5 to 4.0 s after arrow onset was submitted to a CSP
implementation following Ramoser et al. (2000). From the unmixing CSP matrix W (filter x
channels) the channel (column) with the absolute minimum weight and the channel with the
absolute maximum weight were determined, ideally from the first and the last filter (row). If
either the first or the last filter was clearly distorted by artifact, the next filters (up to five from
the front or back respectively) were evaluated. Each of the two bipolar channels consisted of a
signal channel (maximum) and reference channel (minimum). Selected channels had to fulfill
two restrictions. First, each channel should be part of the top/centered 35 channels of the 94-
channel layout, and second, a bipolar channel had to remain within one hemisphere. Midline
electrode positions were considered valid for both hemispheres. Pilot data indicated that this
channel selection procedure returned reasonable results, despite its simplicity. Iterative CSP
procedures as proposed previously (Farquhar et al., 2006; Popescu et al., 2007; He et al.,
2010) were not applied for the channel selection step, as they are not designed to identify
bipolar channels per hemisphere. Figure 2 displays exemplary for right and left hand MI CSP
topographies and the resulting bipolar channels for two individuals, and the group average
maps and frequency maps of the selected electrodes.
--- FIGURE 2 HERE ---
ICA-informed channel selection
For ICA-based electrode selection the EEG raw data were initially filtered between 1 and 40
Hz, segmented into consecutive time intervals of one second and segments containing large
artifacts were rejected. The remaining data were submitted to extended infomax ICA (Bell
and Sejnowski, 1995) as implemented in EEGLAB 10.2.2.4b (Delorme and Makeig, 2004) to
estimate the unmixing weights (W) of 60 independent components (ICs) from the EEG data.
The resulting weights matrix was applied to the raw data (Debener et al., 2010). This strategy
provides a reasonable trade-off between retaining as much variance in the data as possible and
protecting ICA against the negative influence of channel drift and other processes
incompatible with the statistical assumptions of ICA (De Vos et al., 2012). All obtained ICs
were back-projected to the sensor level. Note that this included ICs reflecting stereotypical
artifact, because the goal was to determine the optimal channels for further online application,
in the presence of those stereotypical artifacts. For further analysis data were again filtered
between 8-30 Hz, segmented from 0.5 s before to 4.0 s after arrow stimulus onset and
segments containing large artifacts were rejected.
Bipolar channel selection procedure closely followed the one proposed by Lou et al. (2008),
with few exceptions. In their study involving eight MI trained subjects Lou and colleagues
acquired right and left hand MI data from 32 channels (120 trials per class), in order to reduce
the number of electrodes to two bipolar channel pairs based on ICA. Therefore the authors
first identified one SMR IC for the left and one for the right hemisphere based on three
criteria: (1) the IC activation spectrum had to include a peak in the µ-frequency range (8-13
Hz), (2) the spatial distribution of the source (forward matrix W-1) had to have a focus over
sensorimotor areas; and (3) the 8-30 Hz band-power of the source had to correlate with the
site of the imagination task, that is, it was expected to show a contralateral power suppression
and an ipsilateral enhancement. In the present study the SMR ICs were selected by two
independent evaluators (CZ, SD) and criterion 1 was tightened as in addition a to a first peak
in the µ-frequency range a second one in the β-band had to be present, located approximately
at twice the frequency than the first peak. Moreover, while Lou et al. selected three task-
unrelated α-ICs to represent noise, we used all ICs except for the selected SMR ICs as noise
estimator. By this we could avoid the additional IC selection step. Moreover, we considered
our approach of noise estimation more realistic for the subsequent low-density recordings.
Despite the minor changes regarding the IC selection, during the channel selection procedure
within the defined ICs we closely followed the method proposed by Lou and colleagues.
Specifically, the channel with the absolute maximum weight in the selected SMR IC was
chosen as signal channel. The corresponding reference channel was required (1) to have the
lowest absolute weight in the selected SMR IC; and (2) to have a noise level as similar as
possible to the noise level of the selected signal channel. To this end the signal-to-noise ratio
(SNR), ratio of task-related power (signal) to noise power, was calculated for all channels and
the channel with the highest SNR was chosen as the respective reference channel. The
reference channel for the left signal channel L was calculated as follows:
SNRL(c) = |���������∗�|
|�������∗|�∑ | ����������∗�� |����
Here wcn is an element of W-1, where the first index indicates the channel number and the
second the IC number. Ideally, ICA decomposed the bandpass filtered data in two SMR ICs
(L, R) and 58 noise ICs ni, i=1,…,k. The band power P of the ICs was obtained by bandpass
filtering, squaring and averaging all samples in the 3.5 s time window. Similar to the CSP-
based channel selection approach, channels were required to be part of the top/centered 35
channels, and a bipolar channel had to stay within one hemisphere. Figure 2 displays
exemplary SMR IC topographies and the selected bipolar channels of two subjects for right
and left hand MI, and the normalized group average and relative frequency of the selected
channels for all participants.
Standard channels
The standard bipolar channel pair consisted of close approximations of the 10-20 electrodes
C3 and C4, respectively referenced to FCz.
Classification
OpenViBE was used to calculate the single-trial classification accuracies online as well as
offline. Band power features were computed by first applying a spatial filter. For the first day
CSP was applied to calculate the appropriate coefficients and for days 2 to 4 the three bipolar
channel pairs were used. EEG data were bandpass filtered (8-30 Hz), segmented into motor
imagery intervals covering 0.5 s until 4.0 s after stimulus onset and squared, to obtain the
power. These intervals were subdivided into 40 bins ((3.5-1)/0.0625), resulting in 1 s lasting
time windows with an overlap of 1-0.0625 s. Samples within one time window were
averaged. The resulting features were submitted to linear discriminant analysis where the
mean single-trial classification accuracy was determined with a seven-fold cross-validation
procedure. To this end each dataset was divided randomly into seven parts of equal size, then
feature selection and classifier training was done on 6/7 of the data and obtained predictions
compared to the real labels of the remaining seventh part.
ERD
To compute the ERD% time course, the procedure proposed by Pfurtscheller and Lopes da
Silva (1999) was used, while keeping most parameter settings as for the single-trial
classification in OpenViBE identical. Specifically, to obtain the power samples of a bipolar
channel the difference between reference and signal channel of the filtered signal (8-30 Hz)
was squared. Task related trials were extracted by segmenting the data from -3.0 s before until
4.0 s after stimulus onset. After averaging the power samples across trials, 96 overlapping
time windows were generated and the samples within each time window averaged. ERD%
was computed as follows: ERD%(t)=(A(t)-R)/Rx100, where R is the power in the frequency
band of 2 s baseline period before stimulus onset, and A the power in the frequency band at
time point t, with reference to stimulus onset. By using this equation, ERD% is reflected by
negative numbers. For further analysis, left and right ERD% time courses were averaged and
only contra- and ipsilateral effects are reported, because hemispheric differences were not
considered in this study. For the statistical analysis of ERD% the average of the time window
from 1.0 s to 4.0 s after stimulus onset was used.
To verify that changes in ERD% over the three consecutive days MI practice are caused by
narrow-band modulations, the log power spectra (1-40 Hz) for day 2 and day 4 were
calculated from raw data for the contra- and ipsilateral hemisphere within the MI interval. The
frequency range of 1 to 40 Hz was subdivided into four frequency bands, namely pre-MI
frequency range (1-7 Hz), µ- (8-12 Hz) and β- band (13-30 Hz) and post-MI frequency range
(31-40 Hz).
Iterative CSP
Reducing the number of channels to one bipolar channel pair resulted in a very lean system,
but the leanest system is only in rare cases the one that performs best for a particular
individual. The iterative CSP, as proposed by Popescu et al. (2007) is a channel reduction
approach and can be used to find the optimal tradeoff between the leanest and the best
performing system. The procedure works as follows: Once the unmixing CSP matrix W is
calculated, the least-discriminative channel (the column with the lowest sum of the absolute
values of the first and last filter), is removed and CSP is applied to the remaining N-1
channels. In the present study this process was repeated until only two channels were left. To
do this, Popescu’s approach had to be modified such that not the sum of the first and last two
filters, but the sum of the first and last filter was used to find the least-discriminative channel.
To further explore the relationship between classification accuracy and number as well as
position of channels for every iteration step, the relative importance for any electrode and the
accuracy were determined. Therefore CSP coefficients were calculated on the training
datasets and applied to the two concatenated feedback sessions. Classification accuracy was
again determined using seven-fold cross-validation. Note that the iterative CSP could not be
considered in the first place because it does not result in a bipolar channel pair.
Statistical analysis
All analysis of variance (ANOVA) models consisted of repeated measurement factors. To
analyze if the three sessions recorded within one day differed in classification accuracy, a
one-way ANOVA with factor session (training, feedback 1, feedback 2) was conducted on the
data obtained on the first day with the 94 channel cap. After channel selection an ANOVA
with the factor channel pair (CSP, ICA, standard) was used to statistically test for the
hypothesized higher accuracies for individualized compared to standard channel pairs. To test
for learning over days 2 to 4 and for differences between the channel pairs, classification
accuracies and ERD% as dependent variables were analyzed with two separate ANOVAs,
comprising the factors day (day 2, day 3, day 4) and bipolar channel pair (CSP, ICA,
standard). Furthermore, one ANOVA with factors day (day 2, day 3, day 4) and laterality
(contra-, ipsilateral) was conducted on the ERD% as dependent variable, to statistically check
if any increase in ERD% was caused by changes within the ipsi- or contralateral hemisphere.
To test that changes in ERD% over the practice period are caused by narrow band power
modulations we performed an ANOVA with the factor day (day 2, day 4) and frequency band
(1-7 Hz, 8-12 Hz, 13-30 Hz, 31-40 Hz). To estimate the test-retest reliability of single-subject
measures Pearson correlations were calculated for classification accuracies and ERD%
between day 2 and day 4. KVIQ scores were analyzed by an ANOVA comprising the factors
day (day 1, day 4) and subscale (visual, kinesthetic), to statistically confirm for the assumed
stronger increase in scores for the kinesthetic subtest compared to the visual subtest. For all
ANOVAs, Greenhouse-Geisser epsilon values were used to account for violations of
sphericity. Effect sizes described are η2 (partial eta squared). Paired sample t-tests were
conducted to follow up significant main effects and interaction effects, reported are two-tailed
p-values. Single-subject classification accuracies were considered significantly above chance-
level if they were above a cut-off value. The cut-off value depended on the number of epochs
(two feedback sessions with 40 trials each) and was defined as chance-level (0.5) plus two
standard errors 0.5 + 2 ∗ ��0.5 ∗ 0.5�/80 = 0.612 (Hill and Schölkopf, 2012). Correlation
analyses were performed to explore the possible association of initial KVIQ scores and initial
accuracy, change over the days in KVIQ scores and accuracy as well as change in KVIQ
scores and change in ERD%.
RESULTS
Day 1 – initial online and offline MI accuracy
On the first day, MI single trial accuracy for the training session was on average 84.2%
ranging from 66.7% to 94.8%. This indicated that individuals experienced having control over
their SMR activity, in the absence of extensive training. Online, the single-trial classification
accuracy was 83.1% (range: 64.7%−96.9%) for the first feedback session and 87% for the
second (range: 69.3%−97.7%). Descriptively, higher accuracies for the second feedback
session indicated better control and confirmed that participants remained motivated over the
three recording sessions, but the difference was not significant (F2,30=1.40, p=.263, η²=.09).
Therefore, for all further analyses data were collapsed across the two qualitatively similar
feedback sessions. The single-trial classification accuracies for each individual and group
average results are shown in Figure 3.
After bipolar channel selection the accuracies for every channel pair and all three pairs
together were calculated offline. Note that this was done for the data of the first day. The
single subject and group average classification accuracies in Figure 3 also show the
performance for each individual channel pair (group average for CSP: 71.3% ± 10.1; ICA:
68.3% ± 11; standard: 66.3% ± 10.8) and all three pairs together (group average: 74.4% ±
9.7). Fourteen out of the sixteen participants performed significantly above chance-level when
either all three pairs together, or just the CSP-based channel pair, was considered. This was
the case for ten individuals for the ICA-based and eleven individuals for the standard channel
pair. The three bipolar channel pairs differed significantly in classification accuracy
(F2,30=7.40, p=.003, η2=.33). Pairwise comparisons showed a significant difference between
CSP and standard channel pair (t15=3.39, p=.004) as well as between CSP and ICA channel
pair (t15=2.43, p=.028), indicating that the CSP-informed bipolar channel pair performed best.
No significant difference was found between ICA and standard channel pair (t15=1.69,
p=.112).
--- FIGURE 3 HERE ---
Day 2 to day 4 – learning effects
The classification accuracies over the three days obtained with the different bipolar channel
pairs were analyzed with a two-way ANOVA with factors day (day 2, 3, 4) and channel pair
(CSP, ICA, standard). A significant main effect of day (F2,30=3.62, p=.041, η2=.19) was found
but no significant main effect of channel pair (F2,30=.01, p=.990, η2=.001), suggesting similar
changes in accuracy over days for all bipolar channel pairs. Therefore, in the following
analysis the online accuracies, based on all three channel pairs together, were considered. The
group average for day 2 to 4 of online single trial classification accuracies, based on all three
bipolar channel pairs together, is displayed in Figure 4. Pairwise comparison confirmed that
accuracies were significantly higher on day 4 than on day 2 (t15=-3.58, p=.003) indicating
learning, in terms of increased accuracies over three days practice. Descriptively, accuracies
improved on average from 69.1% on day 2 to 73.3% on day 4. A strong positive correlation
between classification accuracies of the feedback sessions across channel pairs between day 2
and day 4 (r=.86, p<.001) confirmed an excellent test-retest reliability.
--- FIGURE 4 HERE ---
Confirming previous reports MI caused an ERD prominent above the contralateral
sensorimotor area in the MI specific frequency band spanning the 8 to 30 Hz range1. Grand
average ERD% for the days 2, 3 and 4 are displayed in Figure 5 for the contra- and ipsilateral
hemisphere along with the difference between both. Two ANOVAs, one for the contra- and
one for the ipsilateral ERD%, with factors day (day 2, 3, 4) and channel pair (CSP, ICA,
standard) were conducted. Neither for the contralateral nor for the ipsilateral hemisphere a
significant difference on ERD% between channel pairs was found (contralateral: F2,30=2.85,
p=.101, η2=.16; ipsilateral: F2,30=2.80, p=.103, η2=.16), therefore the following analysis was
based on the average of all three channel pairs. Changes in the difference of contra- and
ipsilateral ERD% over three days of MI training were analyzed by the ANOVA with factor
day (day 2, 3, 4), which revealed a significant effect (F2,30=4.64, p=.024, η2=.24). The
difference of contra- and ipsilateral ERD% was significantly higher on day 4 compared to day
2 (t15=2.44, p=.028). This indicates learning in terms of more discriminative ERD% of the
two hemispheres. The relative power improved in the motor imagery period from 11.7% on
day 2 to 17.0% on day 4. As illustrated, the increase in difference between contra- and
ipsilateral ERD% over the three days was related to changes in ERD% of the ipsilateral rather
than of the contralateral hemisphere. An ANOVA with factors laterality (contra-, ipsilateral)
and day (day 2, day 3, day 4) revealed a significant main effect of laterality (F2,30=16.97,
p=.001, η2=.53) and a significant interaction between day and laterality (F2,30=4.64, p=.024,
η2=.24). Within each day the ERD% of the contralateral hemisphere (day 1: -33.2 ± 19.9; day
2: -33.7 ± 18.5; day 3: -33.6 ± 20.6) was significantly higher than that of the ipsilateral
hemisphere (day 1: -21.2 ± 21.7; day 2: -20.5 ± 21.1; day 3: 16.6 ± 24.8) (all p=.001).
Pairwise comparisons on ERD% within the ipsilateral hemisphere revealed a close to
significant difference between day 2 and 4 (t15=-2.10, p=.053), which was not observed for
the contralateral hemisphere (t15=-.29, p=.775). This indicates that the reported changes in
ERD% can be mainly traced back to the ipsilateral hemisphere.
--- FIGURE 5 HERE --- 1 Additional analyses confirmed that no ERD was evident in the pre- (1-7 Hz) and post-frequency of interest (31-
40 Hz) over the three days of MI practice, see Supplementary Figure S1.
In order to test the frequency-band specificity of the training effect, log power spectra were
calculated. Spectra for the two feedback sessions and the three channel pairs were averaged
because the learning effect was restricted to the feedback sessions (cf. figure 4) and no
significant difference between channel pairs was found with regard to the learning effect on
the three consecutive days (cf. figure 5). Figure 6 illustrates the difference of contra- and
ipsilateral hemisphere power at day 2 and day 4 for four frequency bands. An ANOVA with
the factors day (day 2, day 4) and frequency band (1-7 Hz, 8-12 Hz, 13-30 Hz, 31-40 Hz)
revealed a significant main effect of frequency band (F3,45=7.37, p=.002, η²=.33), weather
factor day failed to reach significance (F1,15=4.04, p=.063, η²=.21). Comparing the power
from day 2 and day 4 for each frequency band, a significant increase could be observed for
the MI relevant frequency bands (µ and β), whereas no significant difference was found for
the other frequency bands (1-7 Hz: t15=-0.91, p=.38; 8-12 Hz: t15=-2.39, p=.031; 13-30 Hz:
t15=-2.17, p=.046; 31-40 Hz: t15=-.68, p=.508). This result was confirmed when repeating this
analysis with a higher spectral resolution with a frequency bin width of 2 Hz (see
Supplementary Figure S2). Thus, the learning effect was frequency-band specific rather than
broadband. Excellent test-retest reliabilities were confirmed by a strong positive correlation of
ERD% across channel pairs between day 2 and day 4. These correlations were significant for
the contra- and ipsilateral hemisphere as well as for the difference between both (r=.96,
p<.001; r=.94, p<0.001; r=.90, p<.001).
--- FIGURE 6 HERE ---
KVIQ
To explore the relationship between initial KVIQ scores and initial online accuracy, Pearson
correlations were calculated, but neither for the total nor for the single subscales a significant
correlation (total: r=-.16, p=.54; kinesthetic: r<-.01, p=.99; visual: r=-.33, p=.22) was found.
Furthermore, KVIQ scores from the first and last day were compared in an ANOVA with
factors day (day 1, day 4) and subscale (visual, kinesthetic). A significant main effect for
factor subscale (F1,15=7.23, p=.017, η2=.33) and a marginal day x subscale interaction
(F1,15=4.48, p=.052, η2=.23) was followed up by t-tests. At the first day scores obtained from
the visual subscale were significantly higher than those obtained from the kinesthetic one
(t15=2.69 p=.017). For the kinesthetic subscale, the scores at the last day were significantly
higher than on the first day (t15=-2.49, p=.025), but no significant effect emerged for the
visual subscale (t15=.20 p=.848). Neither for the total test nor for one of the subscales the
changes between the first and the last day in KVIQ scores correlated significantly with the
changes between day 2 and day 4 in accuracy or in ERD%.
Post Hoc Analyses
To explore the relationship between the number of EEG channels and MI accuracy, the data
from day 1 were submitted to an iterative CSP approach. Figure 7 shows the single-trial
classification accuracies for decreasing number of channels, from all 94 down to two
channels. Topographies illustrate at selective iteration stages the relative frequency of non-
eliminated channels, considering all 16 datasets. In most individuals, the number of channels
could be reduced to ten, without dropping dramatically in performance. The most important
ten channels were located on central scalp sites, very similar to the pattern reported by
Popescu et al. (2007). However, individuals differed largely in overall accuracy, as well as in
the impact of channel reduction on MI accuracy. On a single subject level we categorized the
datasets into three groups. A first group (01, 05, 10, 11, 12, 15, 16) delivered a high level
performance with all channels (>82%) and maintained this level until very few channels were
left. A second group (03, 07, 08, 14) performed in the beginning on a lower but still sufficient
level, but the performance worsened with decreasing number of channels and to a different
degree. The performance of individuals in the third group (02, 04, 06, 09, 13) was poor
(<75%) with all 94 channels and remained low with decreasing number of channels.
Topographies illustrating the relative frequencies of the remaining channels across subjects
revealed more clearly electrodes in proximity to sensorimotor cortex with fewer channels. For
the last few iteration stages a pronounced lateralization to the left hemisphere was observable,
presumably reflecting the inclusion of only right-handed participants into this study.
--- FIGURE 7 HERE ---
A further post-hoc analysis evaluated the role of the number of channels for MI performance
on days 2 to 4. This is because the CSP- and ICA-informed channel selection steps were
performed independent from each other. Therefore, in some cases the same channel was
selected twice, which resulted in differences regarding the total number of channels (seven to
ten) that were used for individual subjects days 2 to 4. This might have caused an artificial
increase of the inter-individual differences observed. To evaluate this, individuals were
divided into two groups (group 1: 7 or 8 channels, N = 6; group 2: 9 or 10 channels, N = 10).
A pairwise comparison showed no significant difference in MI accuracy between these two
groups (t5=.31 p=.766).
DISCUSSION
The present study investigated effects of systematic multi-day MI practice in a natural
environment using a portable EEG system. To this end, based on an initial high-density
recording in the laboratory, the number of channels was reduced to a minimum. Two methods
to obtain an individually selected bipolar channel pair were used and compared with a
standard C3/4 versus FCz bipolar pair. The study reports two important findings: Firstly,
individually selected channels outperformed the standard channels, and the pair based on CSP
performed significantly better than the one obtained by ICA. Secondly, with the low-density
individualized cap and the wireless EEG system we found an MI learning effect, expressed in
higher MI accuracies, ERD% and KVIQ scores after three days MI practice in an everyday
environment.
Unfortunately only a limited number of previous studies examined the effect of MI practice
and a direct comparison of accuracies and ERD% across studies is difficult for several
reasons. Firstly, the participants’ level of previous BCI and/or MI experience ranges from
naïve (Blankertz et al., 2008a; Neuper et al., 2009; Friedrich et al., 2013; Ono et al., 2013),
through experienced (Guger et al., 2000; Lou et al., 2008) up to participant selection
depending on previously achieved performances (Wang et al., 2006; 2007b). Furthermore,
different class types (Blankertz et al., 2007; 2008a), number of classes (Friedrich et al, 2013),
instructions (Neuper et al., 2005; Schuster et al., 2011), frequency bands, channel numbers
and applied spatial filters make it even more difficult to compare the performance obtained in
this study with previous findings.
Day 1 – initial online and offline MI accuracy
Taking the described differences between studies into account, the obtained MI accuracies
from the first day using the high-density cap correspond to previous 2-class MI data obtained
from naïve individuals. Experienced participants, however, like those participating in the
study of Lou et al. (2008), can obtain higher accuracies. The difference at the low-density
level, by comparing the accuracies obtained with the bipolar channel pair based on ICA, is
even larger (Lou et al.: mean: 91.1%, range: 69.8% - 98.3%), despite the fact that the present
study closely followed the channel selection procedure by Lou et al. We attribute this
difference to the fact that our participants had no previous MI experience, while those of Lou
et al. were very experienced. Our own results clearly show that having practiced MI for only
three consecutive days is sufficient to improve MI accuracies and ERD%. Since MI practice
leads to more focused SMR topographies (McFarland et al., 2010; Friedrich et al., 2013), the
discrepancy in accuracy between experienced and naïve individuals may be inflated at low-
density levels.
By comparing the three bipolar channel pairs, our results confirmed that individually selected
channel pairs perform better than standard ones. Within the two individually selected bipolar
channel pairs the pair based on the spatial filter CSP significantly outperformed the ICA
based channel pair. Because not only the underlying spatial filter differed, but also the
selection procedures and SMR topographies (unmixing, forward) used were different this
could be due to several reasons. One possibility is that the advantage is inherent in the CSP
algorithm, because the simultaneous variance maximization of the first class and
minimization of the second class provide better-represented features than ICA components.
However, this is not supported by the literature (Naeem et al., 2009; Wang et al., 2012).
Moreover, one could argue that either the channel selection used for determining the CSP
channels and/or the underlying unmixing weights better reflect the mental state of the
participants. Another reason could be that despite careful screening of all ICs by two
independent evaluators, the optimal SMR components were not found. The manual
identification of ICs with the characteristics associated with MI is difficult, especially when
more than one possible SMR component per hemisphere is present. Such a manual
identification of ICs is subjective, time consuming and error prone if not enough expert
knowledge is available. To overcome these limitations, a future goal should be to develop
software that automatically identifies and clusters SMR components, similar to previous
software solutions, which however were not applicable in the present case (Viola et al., 2009;
Viola et al., 2011).
Reducing the number of channels from high-density to bipolar level was motivated through
issues of practicability and affordability. Our results support that by increasing the number of
channels at appropriate sites higher accuracies can be achieved. But how many channels
represent the optimal trade-off between a lean system and acceptable performance is not well
understood. Schröder et al. (2005) already showed that this question should be answered for
each subject individually, as the performance with individual number and individual
positioning of electrodes revealed better accuracies. The iterative CSP approach as proposed
by Popescu et al. (2007) and applied in the present study as a post hoc analysis is one
approach which could be used to identify the optimal number of channels and channel
positions for each individual. Other iterative procedures based on CSP have been proposed
(Farquhar et al., 2006; He et al. 2010), but a comparable procedure for ICA is missing. From
our iterative CSP results one could conclude that, on average, the number of channels
providing a faire trade-off between a lean system with a short set-up time and an acceptable
performance is close to ten. However, the number of channels needed to perform
approximately at the same level as with the full 94-channel cap varied strongly between
individuals. Therefore, the best strategy seems to be to define both, optimal low-density
channel number and optimal layout for each person separately. For excessive longitudinal
training studies the extra effort going into an individualized cap design would pay off quickly.
In addition to the number of electrodes also the type of electrodes used is crucial for a
practicable EEG system. Without doubt applicability of regular MI practice would be
improved further by replacing the traditional wet electrodes with high quality dry electrodes
(Chi et al., 2010; Liao et al., 2012). To the best of our knowledge no previous MI study used
dry electrodes in a home setting and signal quality issues remain with currently available dry
electrodes. Therefore we refrained from using dry electrodes in the present study.
Day 2 to day 4 – learning effects
The most crucial aspect of any rehabilitation program is the resulting benefit for the patient. It
is likely that the patients’ benefit depends very often on the ability to perform a rehabilitation
task appropriately, which means in the case of MI training to imagine moving distinct body
parts. The high relevance of the individual ability to learn MI is reflected among other things
by the attempts of Blankertz et al. (2008a) to show a learning effect within one practice
session. Despite a large number of trials no significant trend in accuracy over time was found,
therefore the authors concluded that one day of MI practice is not enough to show learning
effects. In contrast to this study we kept the recording sessions per day very short (approx. 24
min MI practice per participant per day), in order to sustain attention and keep participants’
compliance up over the four consecutive days MI practice, both particularly important issues
for patients.
Our results show frequency specific MI learning effects over a time range of four consecutive
days. Neuper et al. (1999) similarly observed learning in terms of improved classification
accuracies and increased differences of contra- and ipsilateral ERD% after 5 to 10 sessions in
all of their four experienced individuals with a fixed bipolar channel pair, and Guger et al.
(2000) reported an improvement in classification accuracy for three experienced individuals
after three days of MI practice. In spite of these parallels a direct comparison with our results
is not possible because firstly, Guger and colleagues the actual individual learning effect
could not be obtained due to irregular adaptation of spatial filter coefficients, and secondly,
our participants were novice BCI users. For naïve individuals, improved classification
accuracies and significantly increased ERD% effects were reported by Ono et al. (2013). They
observed that practicing MI over five consecutive days induced prolonged and more robust
contralateral ERD% as well as smaller inter-trial variability compared to the initial session.
The lateralization of the learning effect towards the contralateral side is not supported by our
study however, as our results provide evidence that MI practice leads to a decrease of
ipsilateral rather than an increase in contralateral ERD%. Our observation of an ipsilateral
learning effect could indicate that MI practice goes along with a higher degree of decoupling
of the hemispheres. Testing this idea will require further MEG or high-density EEG studies to
obtain detailed spatial and temporal information of the SMRs (Cheyne, 2013).
More complex 2D motor imagery based BCI experiments also provided evidence that
participants gradually learned to control their SMRs over 21 to 68 sessions (Wolpaw and
McFarland, 2004; McFarland et al., 2010). Royer et al. (2010) reported that their well-trained
participants learned to successfully navigate a helicopter continuously in a 3D space with a 4-
class motor imagery BCI within 8 to 11 sessions. In contrast, some MI studies did not find a
general improvement of performance or larger contra- versus ipsilateral ERD% differences
over several days of practice (Neuper et al., 2009; Friedrich et al., 2013). This discrepancy
may be explained by differences in how frequent the classifier was updated. The importance
of a frequent update of the classifier (Shenoy et al., 2006) and a sufficient number of trials for
stability (Blankertz et al., 2008a) has been reported before, and it is conceivable that
suboptimal settings will hamper MI learning effects. Moreover, in the case of Friedrich et al.
(2013) it should be taken into consideration that a 4-class MI paradigm was used. It is very
likely that learning to operate a 4-class MI paradigm is much more difficult and thus requires
more practice than learning to steer a 2-class paradigm.
To the best of our knowledge the present study is the first to demonstrate MI learning in a
comparatively large sample of naïve participants in an everyday environment. Learning was
reflected in higher classification accuracies and an increased difference of contra- and
ipsilateral ERD% after continued practice in an everyday environment. The learning effect
was likely facilitated by the short duration of the practice sessions, which helped to avoid
exhaustion and boredom. Another probable reason for why we obtained a significant learning
effect despite a relative short practice period is that our practice schedule was well defined,
meaning that every participant practiced on four consecutive days, every day at the same time,
and every day supervised by the same experimenter. The number of sessions was constant
every day and known to the participants. Thus, unlike other studies (Millán et al., 2004;
McFarland et al., 2010; Royer et al., 2010), we avoided varying intervals between practice
days and kept as many variables of the design as possible constant. But in spite of the overall
learning effect, drops in performance between days were evident in nearly all participants.
Such inconsistency is typical for BCI experiments and has been reported previously in studies
conducted in a laboratory environment (McFarland et al., 2010). The drops in performance
likely relate to motivational issues or to attentional lapses. It is important to note that even in
light of inconsistencies in the learning curves, a common trend in learning curves is
recognizable. First, performance improved within one day between sessions. Second, this
improvement increased over days, that is, on the last day the improvement from the first to the
last session was higher than on the first day. This pattern is also evident in the group average
of the present data, and it is in line with observations made by Millán et al. (2004). As a final
observation, our data indicate that the learning effect is limited to the feedback sessions as
performance during the training session was comparable for all four days. This emphasizes
the importance of feedback for MI learning.
With regard to stroke rehabilitation, at present we can only speculate whether patients would
show the same changes in ERD as our healthy volunteers. Neural plasticity after stroke may
compensate for the loss of motor function. On the other hand, unimpaired motor projections
and competitive interaction contribute to maladaptive plasticity, which negatively affects
motor recovery. Therefore one reasonable assumption is that patients show pre-training a
weaker desynchronization of the lesioned hemisphere, and due to maladaptive plasticity a
stronger bilateral activation than healthy individuals (e.g. Johansen-Berg et al, 2002; Grefkes
and Ward, 2013). Training effects could firstly manifest itself in an increase in ERD% in the
hemisphere ipsilateral to the impaired limb, in parallel to the ipsilateral MI learning effect
observed in the present study. This could then be interpreted as minimization of maladaptive
reorganization. The importance of the ipsilateral hemisphere after stroke, as adaptive and
compensatory hemisphere, has been reported before (Fridman et al., 2004; Buch et al., 2008;
Caria et al., 2011). Secondly, we would predict a training-related increase in
desynchronization for the lesioned hemisphere if motor functions of the lesioned hemisphere
recover. Evidently, this prediction can only be investigated in a patient population. We
believe that alongside these speculations our results clearly show that it is important not to
neglect the unlesioned hemisphere in neurorehabilitation training protocols. However, as
bilateral training may prevent learned nonuse but enhance maladaptive plasticity (Takeuchi
and Izumi, 2012), the hands should be unilaterally imagined. Consequently, in the case of 2-
class paradigms, right versus left hand imagination is more advisable than one hand
imagination versus rest.
KVIQ
KVIQ scores obtained on the first day were higher for the visual than for the kinesthetic
subscale, which corresponds to previous findings (Malouin et al., 2007a). Good test-retest
reliabilities of the total KVIQ and its subscales have been reported for non-disabled
individuals (Malouin et al., 2007b) and patients (Malouin et al., 2007b; Randhawa et al.,
2010; Schuster et al., 2012), which allowed for a fair comparison between the KVIQ scores
obtained on the first and last day in our study. Therefore, higher scores for the kinesthetic but
not for the visual subscale on the last day in comparison to the first day indicate improved
kinesthetic motor imagery skills after four days of practice.
Excluded participants
From the initially twenty-two subjects who participated in the study, five had to be excluded
because the obtained bipolar channels pairs were not suitable. For one participant neither the
ICA reveal motor imagery sensitive ICs, nor the CSP result in MI sensitive filter. For four
participants MI ICs could be selected, but for none of them did the CSP result in an MI
sensitive filter, even if the first and last five filters were considered. The mean accuracy for
these four participants using the ICA-based channel pair was 53.3%, indicating that their
EEGs were not characterized by clear sensorimotor cortex activity. This may explain why the
CSP approach failed. The ICA approach seems less affected by this lack of clear sensorimotor
cortex activity, which is also reflected in the finding that with ICA SMR components can
even be obtained from resting state EEG (Wang et al., 2012). But regardless of spatial filter,
the inability of a non-negligible proportion of individuals to successfully do MI has been
reported in many previous studies (Blankertz et al. 2010). Guger et al. (2003) for instance
tested 99 naïve participants and found that only approximately half of them reached
classification accuracies above 70%. To improve the applicability of MI based neurofeedback
further studies are clearly required that investigate the relationship between MI BCI
performance and individual differences in the neurophysiological correlates of MI.
Future issues
To conclude, this study demonstrates that MI performance and ERD% can be significantly
improved after three days MI practice in an everyday environment using individually
designed electrode caps in combination with a wireless EEG system. Furthermore, our results
show that MI neurofeedback can be realized with a bipolar channel pair and better so by
taking individually-selected sites instead of standard channel sites. If taking this finding a step
further then the system best adapted to a user should return the highest performance. That is,
performance can likely be maximized when in addition to the individual selection of channel
number and position, the number of features and the features themselves (Schröder et al.,
2003; Fabiani et al., 2004), as well as the frequency range (Blankertz et al., 2006; 2007;
2008a) are determined individually. It would be desirable to implement this in an automatic
and user-friendly way, similar to recent advances in P300 speller application (Kaufmann et
al., 2012).
Moreover, our results show that MI practice leads to an increase in the difference between
contra- and ipsilateral ERD%. Others observed changes in SMR topographies following MI
practice (McFarland, 2010; Friedrich et al., 2013). It is now important to evaluate how
changes in SMR relate to clinical improvement, for instance after stroke (Buch et al., 2008).
But in the meanwhile tentative recommendations for long-term MI based neurorehabilitation
protocols can be derived. To achieve optimal therapeutic success it seems important to collect
high-density reference EEG data after regular intervals of MI practice and, if necessary, to
adapt individual low-density channel positions.
ACKNOWLEDGEMENTS
We thank Jeremy Thorne and Martin Bleichner for proofreading the manuscript. CZ was
funded by the PhD program ‘Signals and Cognition’ (Niedersächsisches Ministerium für
Wissenschaft und Kultur). CK is supported by grant KR 3433/2-1, German Research
Foundation (DFG).
LEGEND TO FIGURES
Fig. 1 Experimental paradigm and pictures illustrate the different recording environments and
EEG systems. Left: Rear view of a participant wearing the high-density EEG system while
performing the motor imagery task in the laboratory, as on the first day. Right: Participant
performing the same task in an everyday environment using a low-density wireless EEG
system, as on days 2 to 4. The amplifier is fixed to the lower edge of the EEG cap. Electrodes
are placed forming two bipolar channel pairs (highlighted in red: ICA-based, highlighted in
green: CSP-based) in accordance to the scalp distribution of the sensory motor rhythm of the
participant.
Fig. 2 CSP (a, b) and ICA (e, f) SMR topographies for right (a,e) and left (b,f) hand MI of two
exemplary subjects and the normalized grand average from N = 16 subjects (bottom row).
Middle columns (c, d) displays the selected bipolar channel pairs respectively and the relative
frequency of selected channels (bottom row; blue = reference, orange = signal channel).
Fig. 3 Single subject and group average EEG single-trial classification results plotted as
percent accuracy for: online high-density performance using CSP (blue) and offline calculated
low-density performance for all channel pairs together (grey), CSP (green), ICA (red) and
standard channel pair (black).
Fig. 4 Group average (N=16) EEG single-trial classification results of all three channel pairs
together plotted as percent accuracy for the training, feedback 1 and feedback 2 sessions for
days 2, 3 and 4.
Fig. 5 Grand average ERD% from N=16 for every bipolar channel. Straight line indicate
contralateral and dotted line ipsilateral bipolar channel, with respect to stimuli direction (a).
Differences between contra- and ipsilateral ERD% for all three channel pairs, horizontal line
represents the average ERD% for the MI time window (b). Grey shaded area indicate time
window of stimulus onset.
Fig. 6 Log power spectra of four frequency bands for days 2 and 4. Shown is the difference
between contra- and ipsilateral hemisphere, with respect to stimulus direction. The difference
is significantly larger on day 4 in the frequency bands covering the µ (8-12 Hz) and β (13-30
Hz) range.
Fig. 7 Single subject and group average offline classification accuracies for decreasing
number of channels, ranging from all 94 (left) to 2 (right side). Topographies show the
relative frequency of remaining channels at selective iteration stages considering all subjects.
REFERENCES
Bell AJ, Sejnowski TJ. An information-maximization approach to blind source separation and
blind deconvolution. Neural Comput 1995;7:1129-1159.
Billows IJ, Khan S, Sterr A. Exploring sleep, sleepiness and fatigue in chronic stroke: A mixed methods study. Int J Stroke 2013;8:73-73.
Blankertz B, Dornhege G, Krauledat M, Müller KR, Curio G. The non-invasive Berlin brain–
computer interface: fast acquisition of effective performance in untrained subjects.
Neuroimage 2007;37:539-550.
Blankertz B, Dornhege G, Krauledat M, Müller KR, Kunzmann V, Losch F, et al. The Berlin
Brain-Computer Interface: EEG-based communication without subject training. IEEE
Trans Neural Syst Rehab Eng 2006;14:147-152.
Blankertz B, Losch F, Krauledat M, Dornhege G, Curio G, Müller KR. The Berlin Brain–Computer Interface: accurate performance from first-session in BCI-naive subjects.
IEEE Trans Biomed Eng 2008a;55:2452-2462.
Blankertz B, Sannelli C, Halder S, Hammer EM, Kübler A, Müller KR, et al.
Neurophysiological predictor of SMR-based BCI performance. Neuroimage
2010;51:1303-1309.
Blankertz B, Tomioka R, Lemm S, Kawanabe M, Müller KR. Optimizing spatial filters for robust EEG single-trial analysis. IEEE Signal Processing Mag 2008b;25:41-56.
Buch E, Weber C, Cohen LG, Braun C, Dimyan MA, Ard T, et al. Think to Move: a
Neuromagnetic Brain-Computer Interface (BCI) System for Chronic Stroke. Stroke
2008;39:910-917.
Caria A, Weber C, Brötz D, Ramos A, Ticini LF, Gharabaghi A, et al. Chronic stroke
recovery after combined BCI training and physiotherapy: A case report.
Psychophysiology 2011;48:578-582.
Cheyne D. MEG studies of sensorimotor rhythms: A review. Exp Neurol 2013;245:27-39.
Chi YM, Jung TP, Cauwenberghs G. Dry-contact and noncontact biopotential electrodes:
methodological review. IEEE Rev Biomed Eng 2010;3:106-19.
Cincotti F, Pichiorri F, Aricò P, Aloise F, Leotta F, De Vico Fallani F, et al. EEG-based
Brain-Computer Interface to support post-stroke motor rehabilitation of the upper limb.
Conference Proceedings - IEEE Eng Med Biol Mag 2012:4112-4115.
De Vos M, Thorne JD, Yovel G, Debener S. Let’s face it, from trial to trial: Comparing procedures for N170 single-trial estimation. Neuroimage 2012;63:1196-1202.
De Vos M, Gandras K, Debener S. Towards a truly mobile auditory brain-computer interface:
Exploring the P300 to take away. Psychophysiology 2013;91:46-53.
De Vos M, Kroesen M, Emkes R, Debener S. P300 speller BCI with a mobile EEG system: Comparison to a traditional amplifier. J Neural Eng. doi: 10.1088/1741-
2560/11/3/036008.
Debener S, Minow F, Emkes R, Gandras K, De Vos M. How about taking a low-cost, small,
and wireless EEG for a walk? Psychophysiology 2012;49:1449-1453.
Debener S, Thorne JD, Schneider TR, Viola FC. Using ICA for the Analysis of Multi-
Channel EEG Data. In: Ullsperger M, Debener S, editors. Simultaneous EEG and fMRI:
Recording, Analysis, and Application. Oxford University Press, 2010:121-133.
Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG
dynamics including independent component analysis. J Neurosci Methods 2004;134:9-
21.
Delorme A, Mullen T, Kothe C, Akalin AZ, Bigdely-Shamlo N, Vankov A, et al. EEGLAB, SIFT, NFT, BCILAB, and ERICA: new tools for advanced EEG processing. Comput
Intell Neurosci 2011;130714. doi:10.1155/2011/130714.
Fabiani GE, McFarland DJ, Wolpaw JR, Pfurtscheller G. Conversion of EEG activity into
cursor movement by a brain-computer interface (BCI). IEEE Trans Neural Syst Rehab
Eng 2004;12:331-338.
Farquhar J, Hill J, Lal TN, Schölkopf B. Regularised CSP for sensor selection in BCI. In:
Müller-Putz GR, Brunner C, Scherer R, Schlögl A, Wriessnegger S, Pfurtscheller G,
editors. Proceedings of the 3rd International Brain-Computer Interface Workshop and
Training Course 2006. Graz, 2006:14-15.
Fridman EA, Hanakawa T, Chung M, Hummel F, Leiguarda RC, Cohen LG. Reorganization of the human ipsilesional premotor cortex after stroke. Brain 2004;127:747-758.
Friedrich EVC, Scherer R, Neuper C. Long-term evaluation of a 4-class imagery-based brain–
computer interface. Clin Neurophysiol 2013;124:916-927.
Grefkes C, Ward NS. Cortical reorganization after stroke: how much and how functional? Neuroscientist 2014;20:56-70.
Guger C, Edlinger G, Harkam W, Niedermayer I, Pfurtscheller G. How many people are able
to operate an EEG-based brain-computer interface (BCI)? IEEE Trans Neural Syst
Rehab Eng 2003;11:145-147.
Guger C, Ramoser H, Pfurtscheller G. Real-time EEG analysis with subject-specific spatial
patterns for a brain-computer interface (BCI). IEEE Trans Rehabil Eng 2000;8:447-456.
He L, Gu Z, Li Y, Yu Z. Classifying Motor Imagery EEG Signals by Iterative Channel
Elimination according to Compound Weight. In: Wang FL, Deng H, Gao Y, Lei J, editors. Artificial Intelligence and Computational Intelligence, Lecture Notes in
Computer Science. Berlin Heidelberg: Springer, 2010:71-78.
Hill NJ, Schölkopf B. An online brain–computer interface based on shifting attention to
concurrent streams of auditory stimuli. J Neural Eng 2012;9. 026011.
Holz EM, Botrel L, Kübler A. Bridging Gaps: Long-Term Independent BCI Home-Use by a
Locked-In End-User. In: Proceedings of the TOBI Workshop IV, Sion, 2013:35-36.
Jasper H, Penfield W. Electrocorticograms in man: Effect of voluntary movement upon the
electrical activity of the precentral gyrus. Archiv für Psychiatrie und Zeitschrift Neurologie 1949;183:163-174.
Jeannerod M. Mental Imagery in the Motor Cortex. Neuropsychologia 1995;33:1419-1432.
Jeannerod M. Neural Stimulation of Action: A Unifying Mechanism for Motor Cognition.
Neuroimage 2001;14:103-109.
Johansen-Berg H, Rushworth MF, Bogdanovic MD, Kischka U, Wimalaratna S, Matthews
PM. The role of ipsilateral premotor cortex in hand movement after stroke. PNAS
2002;99:14518-23.
Kaufmann T, Völker S, Gunesch L, Kübler A. Spelling is Just a Click Away - A User-
Centered Brain-Computer Interface Including Auto-Calibration and Predictive Text Entry. Front Neurosci 2012;6:72.
Kriegeskorte N, Simmons WK, Bellgowan PS, Baker CI. Circular analysis in systems
neuroscience: the danger of double dipping. Nat Neurosci 2009;12:535-40.
Kübler A, Nijboer F, Mellinger J, Vaughan TM, Pawelzik H, Schalk G, et al. Patients with
ALS can use sensorimotor rhythms to operate a brain-computer interface. Neurology 2005; 64:1775-1777.
Langhorne P, Coupar F, Pollock A. Motor recovery after stroke: a systematic review. Lancet
Neurol 2009;8:741-54.
Liao LD, Chen CY, Wang IJ, Chen SF, Li SY, Chen B, et al. Gaming control using a wearable and wireless EEG- based brain-computer interface device with novel dry
foam-based sensors. J Neuroeng Rehabil 2012; 9:5.
Lou B, Hong B, Gao X, Gao S. Bipolar electrode selection for a motor imagery based brain–
computer interface. J Neural Eng 2008;5:342-349.
Malouin F, Richards CL, Durand A, Doyon J. Clinical Assessment of Motor Imagery After
Stroke. Neurorehabil Neural Repair 2007a;22:330-340.
doi:10.1177/1545968307313499.
Malouin F, Richards CL, Jackson PL, Lafleur MF, Durand A, Doyon J. The Kinesthetic and Visual Imagery Questionnaire (KVIQ) for Assessing Motor Imagery in Persons with
Physical Disabilities: A Reliability and Construct Validity Study. J Neurol Phys Ther 2007b;31:20-29.
McFarland DJ, Miner LA, Vaughan TM, Wolpaw JR. Mu and beta rhythm topographies
during motor imagery and actual movements. Brain Topogr 2000;12:177-186.
McFarland DJ, Sarnacki WA, Wolpaw JR. Electroencephalographic (EEG) control of three-
dimensional movement. J Neural Eng 2010;7:252-259.
Millán JR, Renkens F, Mouriño J, Gerstner W. Brain-Actuated Interaction. Artificial Intelligence 2004;159:241-259.
Müller-Gerking J, Pfurtscheller G, Flyvbjerg H. Designing optimal spatial filters for single-
trial EEG classification in a movement task. Electroenceph clin Neurophysiol
1999;110:787-798.
Naeem M, Brunner C, Pfurtscheller G. Dimensionality Reduction and Channel Selection of
Motor Imagery Electroencephalographic Data. Comput Intell Neurosci 2009;1:1-8.
Neuper C, Scherer R, Reiner M, Pfurtscheller G. Imagery of motor actions: Differential
effects of kinesthetic and visual–motor mode of imagery in single-trial EEG. Cognitive Brain Res 2005;25:668-677.
Neuper C, Scherer R, Wriessnegger S, Pfurtscheller G. Motor imagery and action
observation: Modulation of sensorimotor brain rhythms during mental control of a
brain–computer interface. Clin Neurophysiol 2009;120:239-247.
Neuper C, Schlögl A, Pfurtscheller G. Enhancement of left-right sensorimotor EEG
differences during feedback-regulated motor imagery. Electroenceph cin Neurophysiol 1999;16:373-382.
Oldfield RC. The assessment and analysis of handedness: The Edinburgh inventory.
Neuropsychologia 1971;9:97-113.
Ono T, Kimura A, Ushiba J. Daily training with realistic visual feedback improves
reproducibility of event-related desynchronisation following hand motor imagery. Clin Neurophysiol 2013;124:1779-1786.
Ortner R, Irimia DC, Scharinger J, Guger C. A motor imagery based brain-computer interface
for stroke rehabilitation. Stud Health Technol Inform 2012;181:319-323.
Pfurtscheller G, Aranibar A. Evaluation of event-related desynchronization (ERD) preceding
and following voluntary self-paced movement. Electroenceph clin Neurophysiol 1979;46:138-146.
Pfurtscheller G, Brunner C, Schlögl A, Lopes da Silva FH. Mu rhythm (de)synchronization
and EEG single-trial classification of different motor imagery tasks. Neuroimage
2006;31:153-159.
Pfurtscheller G, Lopes da Silva FH. Event-related EEG/MEG synchronization and
desynchronization: basic principles. Electroenceph clin Neurophysiol 1999;110:1842-
1857.
Pfurtscheller G, Neuper C. Motor imagery and direct brain-computer communication. Proc IEEE 2001;89:1123-1134.
Pfurtscheller G, Neuper C, Flotzinger D, Pregenzer M. EEG-based discrimination between
imagination of right and left hand movement. Electroenceph clin Neurophysiol
1997;103:642-651.
Popescu F, Fazli S, Badower Y, Blankertz B, Müller KR. Single Trial Classification of Motor
Imagination Using 6 Dry EEG Electrodes. PLoS ONE 2007;2:e637.
Prasad G, Herman P, Coyle D, McDonough S, Crosbie J. Applying a brain-computer interface
to support motor imagery practice in people with stroke for upper limb recovery: a
feasibility study. J Neuroeng Rehabil 2010;7:60.
Ramoser H, Müller-Gerking J, Pfurtscheller G. Optimal spatial filtering of single trial EEG
during imagined hand movement. IEEE Trans Rehabil Eng 2000;8:441-446.
Randhawa B, Harris S, Boyd LA. The Kinesthetic and Visual Imagery Questionnaire Is a Reliable Tool for Individuals With Parkinson Disease. J Neurol Phys Ther 2010;34:161-
167.
Renard Y, Lotte F, Gibert G, Congedo M, Maby E, Delannoy V, et al. OpenViBE: an open-
source software platform to design, test, and use brain-computer interfaces in real and
virtual environments. Presence 2010;19:35-53.
Royer AS, Doud AJ, Rose ML, He B. EEG control of a virtual helicopter in 3-dimensional
space using intelligent control strategies. IEEE Trans Neural Syst Rehab Eng
2010;18:581-589.
Schröder M, Bogdan M, Hinterberger T, Birbaumer N. Automated EEG feature selection for
brain computer interfaces. In: Proceedings of the 1st International IEEE EMBS
Conference on Neural Engineeging, Capri Island, 2003:626-629.
Schröder M, Lal TN, Hinterberger T, Bogdan M, Hill NJ, Birbaumer N, et al. Robust EEG
channel selection across subjects for brain-computer interfaces. EURASIP JASP 2005
2005:3103-3112.
Schuster C, Hilfiker R, Amft O, Scheidhauer A, Andrews B, Butler J, et al. Best practice for
motor imagery: a systematic literature review on motor imagery training elements in
five different disciplines. BMC Med 2011;9:75. http://www.biomedcentral.com/1741-
7015/9/75.
Schuster C, Lussi A, Wirth B, Ettlin T. Two assessments to evaluate imagery ability:
translation, test-retest reliability and concurrent validity of the German KVIQ and
Imaprax. BMC Med Res Methodol 2012;12:127. http://www.biomedcentral.com/1471-
2288/12/127.
Sellers EW, Vaughan TM, Wolpaw JR. A brain-computer interface for long-term independent
home use. Amyotroph Lateral Scler 2010;11:449-455.
Shenoy P, Krauledat M, Blankertz B, Rao RPN, Müller KR. Towards adaptive classification
for BCI. J Neural Eng 2006;3. doi:10.1088/1741-2560/3/1/R02.
Takeuchi N, Izumi S. Maladaptive plasticity for motor recovery after stroke: mechanisms and
approaches. Neural Plast 2012;2012:359728.
Viola FC, Thorne JD, Bleeck S, Eyles J, Debener S. Uncovering auditory evoked potentials
from cochlear implant users with independent component analysis. Psychophysiology
2011;48:1470-1480.
Viola FC, Thorne JD, Edmonds B, Schneider T, Eichele T, Debener S. Semi-automatic
identification of independent components representing EEG artifact. Clin Neurophysiol
2009;120:868-877.
Wang Y, Hong B, Gao X, Gao S. Phase synchrony measurement in motor cortex for
classifying single-trial EEG during motor imagery. In: Engineering in Medicine and Biology Society 2006 - 28th Annual International Conference of the IEEE, New York,
2006:75-78.
Wang Y, Hong B, Gao X, Gao S. Design of electrode layout for motor imagery based brain–
computer interface. Electron Lett 2007a;43:557-558.
Wang Y, Hong B, Gao X, Gao S. Implementation of a brain-computer interface based on
three states of motor imagery. In: Engineering in Medicine and Biology Society, 2007 -
29th Annual International Conference of the IEEE, Lyon, 2007b:5059-5062.
Wang Y, Wang YT, Jung TP. Translation of EEG Spatial Filters from Resting to Motor Imagery Using Independent Component Analysis. PLoS ONE 2012;7:e37665.
Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G, Vaughan TM. Brain–computer
interfaces for communication and control. Clin Neurophysiol 2002;113:767-791.
Wolpaw JR, McFarland DJ. Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans. Proc Natl Acad Sci USA 2004;101:17849-17854.