Hindawi Publishing Corporation
Computational Intelligence and Neuroscience
Volume 2009, Article ID 104180, 6 pages
Discriminationof Motor Imagery-InducedEEGPatterns in
Patients withCompleteSpinalCord Injury
G.Pfurtscheller,1P.Linortner,1R.Winkler,2G.Korisek,2and G.M¨ uller-Putz1
1Laboratory of Brain-Computer Interfaces, Institute for Knowledge Discovery, Graz University of Technology,
Krenngasse 37, 8010 Graz, Austria
2Rehabilitation Clinic Tobelbad, Krenngasse 37, Dr.-Georg-Neubauer-Straße 6, 144 Tobelbad, Austria
Correspondence should be addressed to G. Pfurtscheller, email@example.com
Received 30 October 2008; Accepted 11 February 2009
Recommended by Fabio Babiloni
EEG-based discrimination between different motor imagery states has been subject of a number of studies in healthy subjects. We
detail we studied pair-wise discrimination functions between the 3 types of motor imagery. The following classification accuracies
(mean ± SD) were obtained: left versus right hand 65.03% ± 8.52, left hand versus feet 68.19% ± 11.08, and right hand versus
feet 65.05% ± 9.25. In 5 out of 8 paralegic patients, the discrimination accuracy was greater than 70% but in only 1 out of 7
tetraplagic patients. The present findings provide evidence that in the majority of paraplegic patients an EEG-based BCI could
achieve satisfied results. In tetraplegic patients, however, it is expected that extensive training-sessions are necessary to achieve a
good BCI performance at least in some subjects.
Copyright © 2009 G. Pfurtscheller et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
Functional magnetic resonance imaging (fMRI) and EEG
studies have shown that executed and imagined movement
activates overlapping and/or similar neural networks in
primary motor and related areas [1–3]. This equivalence of
motor execution and motor imagery in relation to cortical
activation is one prerequisite for the restoration of motor
functions in para- and/or tetraplegic patients using a brain-
computer interface (BCI; [4, 5]). Whether patients with
complete spinal cord injury (SCI) are able to control their
brain oscillations reliable and safe through imagined limb
movements and operate herewith a BCI is however still an
Sensorimotor rhythms such as mu and central beta
oscillations can be modified by executed and imagined
movement [6–10]. By using multichannel EEG recordings
and applying pair-wise discrimination functions to the EEG
signals it is possible to discriminate between 3 different types
of motor imagery (right or left hand or foot) [11, 12]. In this
study we addressed the following questions: (i) is it possible
to discriminate pair-wise between 3 motor imagery-related
complete spinal cord injury and (ii) is this discrimination
different for paraplegic and tetraplegic patients. When a
reliable detection of imagery-related brain states in ongoing
EEG is possible the BCI output signal can be used to control,
for example, a neuroprosthesis .
2.1. Subjects and Experimental Task. The patient group
consisted of 15 patients (four females and eleven males) aged
from 16 to 64 years (M = 41 years, SD = 14.50). All patients
suffered from a complete sensor and motor paralysis at ASIA
32.9 years prior to the measurements. Seven patients were
tetraplegic, and eight patients were paraplegic. Information
on the patients is summarized in Table 1.
Measurements were carried out at the Rehabilitation Cli-
nic Tobelbad (Austria). The experiment was divided into 6–
8 runs (depending on the physical condition of the patient),
each consisting of 30 trials of three different motor imagery
tasks (10 trials each). Between those runs participants could
2 Computational Intelligence and Neuroscience
Table 1: Patients characteristics.
Date of birth (year)
Date injury (year)
Number of trials (artifact-free/total)
0.5 ··· 2.5(s)
Figure 1: (a) Electrode positions. (b) Timing and experimental paradigm.
(and were encouraged to) take short breaks for recovery and
in order to avoid fatigue.
Each trial began with the presentation of a fixation cross
at the centre of the monitor, followed by a short warning
tone at second 2. At second 3, an arrow pointing randomly
resp.), appeared on the screen for 1.25seconds, additionally
to the fixation cross. The fixation cross remained displayed
on the screen until the end of the trial at second 8, indicating
that the imagination still had to be performed. This implies a
motor imagery lasting for 5seconds was required. After that,
a blank screen was presented until the beginning of the next
trial. This intertrial period varied randomly between 0.5 and
Timing and experimental paradigm are displayed in
2.2. EEG Recording. Continuous EEG signals were recorded
from a grid of fifteen sintered Ag/AgCl ring electrodes
(Easycap, Germany) that were mounted orthogonally in
both, horizontal and vertical directions, over the electrode
positions C3, Cz, and C4 (according to the international
10–20 electrode system, cf. Figure 1(a)). The closely spaced
interelectrode distance was 2.5cm. All electrodes were refer-
Computational Intelligence and Neuroscience3
Table 2: Classification accuracy (%) of the maximal peak and its latency (delay) after cue onset for all 15 patients and all combinations.
Accuracies in bold differ significantly from chance level according to the number of trials (cf. ).
Left versus right hand
Left hand versus feet
Right hand versus feet
Figure 2: Discrimination time courses for a length of 5s after cue onset. The onset of cue presentation is at second 3. Data from all 15
patients and all 3 brain states are displayed: right versus left hand MI (left panel), left hand versus feet MI (middle panel), and right hand
versus feet MI (right panel).
at the right mastoid. Impedances were kept below 5kOhm.
For monopolar EEG derivation a portable amplifier (g.tec,
Graz, Austria) was used. Signals were digitized at 256Hz and
bandpass filtered between 0.5 and 100Hz. Sensitivity was set
to 100μV and a notch filter at 50Hz was used.
2.3. Data Analysis. The method of Common Spatial Patterns
(CSP) and Fischer’s linear discriminant analysis (LDA)
classifier were used to discriminate between any 2 classes.
dimensional spatial subspace in such a way that the variances
of the filtered time series are optimal for discrimination.
The projection matrix, consisting of the weights of the EEG
channels, is sorted in descending order of the eigenvalues.
Before applying CSP and LDA, a fully automated method for
reducing EOG artifacts was applied on the data. Then, the
EEG recordings were visually inspected for remaining EOG
and EMG artifacts and filtered between 8–30Hz. To get a
procedure was adopted. The EEG data from each trial was
divided into time segments of 1s overlapping by half of their
length. For further details see [11, 14].
4 Computational Intelligence and Neuroscience
Left versus right
Left hand versus
Right hand versus
Figure 3: Two-class classification accuracy for paraplegic and
2.4. Calculation of Time-Frequency Maps. To enhance local
oscillations, orthogonal source derivations (Laplacian) were
calculated . After triggering the data, trials of 10s
duration were obtained including 3seconds before the cue.
The quantification of ERD/ERS was carried out in four steps:
band pass filtering of each trial, squaring of samples (with
smoothing) and subsequent averaging over trials and over
sample points. The ERD/ERS is defined as the percentage
power decrease (ERD) or power increase (ERS) in relation
to a one-second reference interval (0.5–1.5seconds ) before
the warning tone . ERD/ERS values corresponding to 2-
Hz frequency bands ranging from 6–18Hz (with an overlap
of 1Hz) and 4-Hz frequency bands ranging from 18–38Hz
(with an overlap of 2Hz) were calculated. All values for
one EEG channel were subsequently used to construct time-
bootstrap statistic to calculate confidence intervals with
α = 0.05.
2.5. Statistical Analysis. An ANOVA was computed in order
to examine whether paraplegic versus tetraplegic patients
differregarding reachedclassification accuracy.This ANOVA
consisted of the between-subject variable SCI (2 levels:
paraplegics and tetraplegics) and the within-subject variable
ACCURACY (3 levels: left hand versus right hand, left hand
versus feet and right hand versus feet).
The power of discrimination between two different brain
states is indicated by the classification accuracy of single EEG
trials analysed within 1-second time windows. The discrimi-
nation time courses for epochs of 6seconds (with 1second
Table 3: Mean classification accuracy for tetraplegic and paraplegic
Spinal cord injury
L versus R (%)
L versus F (%)
R versus F (%)
prior to cue-onset) for all task combinations (right versus
left hand, left hand versus feet, and right hand versus feet)
are shown in Figure 2. The maximal classification accuracies
of the first peak together with the corresponding latencies,
measured from cue onset are summarized in Table 2. The
mean accuracy of all subjects (±SD) was 65.03% ± 8.51 (left
versus right hand MI), 68.18% ± 11.08 (left hand versus
feet MI) , and 65.05% ± 9.25 (right hand versus feet MI),
respectively. (See Table 3 for the mean accuracy of paraplegic
versus tetraplegic patients.)
In the tetraplegic patient group only one out of seven
tetraplegics had an accuracy >70% while from the para-
plegics five out of 8 reached a classification accuracy >70%.
An accuracy of 70% is the border, where control can be
possible . In the majority of participants, feet motor
imagery was involved in the best discrimination between two
brain states (see Figure 2 and Table 2).
The results of the ANOVA show that the main effect
ACCURACY is insignificant. Paraplegic patients (M =
69.3%) do not differ from tetraplegic patients (M = 62.41%),
F(1,13) = 530.292, p = .151. Furthermore, no significant
effect emerges for the three classification accuracies, left (L)
hand versus right (R) hand (M = 64.85%), left (L) hand
versus feet (F) (M = 67.89%), and right (R) hand versus feet
(F) (M = 64.84%), but a tendency can be seen, F(2,26) =
2.877, p =.074 (cf. Figure 3).
Although the discriminations of any two different brain
over premotor, motor, and parietal areas, different patterns
were found in spatially filtered (Laplacian) recordings over
the primary motor areas (electrode positions C3, Cz, and
C4). For illustration, time-frequency maps (ERD-maps) of
two representative subjects are displayed in Figure 4. In
subject P02 (Figure 4(a)) clearly visible is the beta increase
(ERS) at Cz during hand MI and the beta decrease (ERD)
at Cz during feet MI. No clear EEG reactivity patterns
can be recognized in subject P01 (Figure 4(b)). In P02
a high classification accuracy was obtained while in P01
no discrimination between the motor imagery states was
In our study, we applied a classification procedure to
multichannel, single-trial EEG data recorded during classical
brain-computer interface training sessions with 3-motor
Computational Intelligence and Neuroscience5
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
C3CzC4 C3CzC4 C3CzC4
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 S
C3CzC4C3 CzC4C3 Cz C4
Figure 4: Time-frequency-maps for the three types of motor imagery (left hand, right hand, and feet) computed at electrode positions C3,
CZ, C4 (Laplacian) exemplarily for (a) a patient with good performance (p02) and (b) a patient with bad performance (p01).
18]. One method suitable for studying temporal aspects of
brain activation using multichannel EEG recordings consists
in computing common spatial patterns (CSPs) . This
CSP-method leads to spatial filters that are optimal in the
sense that they extract signals which maximally discriminate
between any 2 conditions. A subsequent linear classification
of these extracted signals results in a good recognition rate.
With the CSP-method it is possible to study the separability
of EEG patterns associated with 2-motor imagery states with
a high time resolution.
The discrimination time courses in the patients with
shape and magnitude and started in generally with an initial
peak about 1.5seconds after cue-onset, with a fast increase
before and a slow decline thereafter (Figure 2). The great
intersubject-variability may be explained by the used mental
strategy (e.g., visual versus kinaesthetic motor imagery,
), the vividness of the imagery process, the mental effort
and other psychological factors as, for example, motivation
and attention. Even in one and the same subject the same
mental motor imagery strategy can result in completely
different EEG reactivity patterns dependent on the degree of
imagined effort .
The main finding of the present study is that there
are distinct EEG patterns in the majority of patients with
complete spinal cord injury when they imagine different
movements of hands and feet the first time. These patterns
rate was relatively low around 67%. In contrast, motor
imagery in healthy subjects results in clearly discriminable
EEG patterns, when 2-motor imagery tasks are compared.
Blankertz et al.  reported a mean classification accuracy
of 88.4% in a so-called calibration session with 3 types of
motor imagery (right hand, left hand, and right foot) in
untrained healthy subjects. This data are based on 128 EEG
channels and CSP analysis. Also with CSP analysis applied
to 32 EEG channels mean classification accuracies between
80.0% and 83.3% are reported for left versus right hand MI
and hand versus feet MI . In both studies in the majority
of subjects the best classification results were achieved when
foot MI was involved. One major difference between healthy
subjects and patients is very often that patients have very
often cramps and/or spasms and therefore a number of
muscle artefacts in the EEG (see e.g., Table 1 artefact-free
versus total trials).
Of interest is a recently published fMRI study where
control subjects and patients had to kinaesthetically imagine
movements of their feet . In the paraplegic patient group
the primary motor cortex was consistently activated, even
to the same degree as during movement execution in the
healthy controls. In contrast to this one other study 
reported inconsistent fMRI activation in the primary motor
patients. Of interest is that in the study of Alkadhi et al. ,
scores of motor imagery in paraplegics and the activation
(fMRI BOLD signal) in cortical areas including the primary
motor cortex and the supplementary motor area (SMA).
This can be interpreted that vividness of motor imagery
and/or their mental effort plays an important role in cortical
activation and is perhaps more intensive in SCI patients than
in healthy controls.
One point needs discussion, namely, the slightly higher
(but not significant) classification accuracy of hand versus
6 Computational Intelligence and Neuroscience
feet MI as compared to right versus left hand MI found
in patients but also reported in healthy subjects. This can
be interpreted to mean that the EEG patterns induced by
feet or foot MI are better discriminable from the brain state
associated with either left or right hand MI. One reason
for this could be the antagonistic behaviour of the upper
mu ERD and ERS during motor imagery known as “focal
ERD/surround ERS” . Feet MI results not only in a
midcentrally focused mu and/or beta ERD but very often
also in a bilateral mu ERS over the hand representation area
. These authors reported on a much larger difference in
band power changes in the 10–12Hz frequency band when
different (hand versus foot MI) and not homologous limbs
(right versus left hand MI) are compared.
In conclusion, we demonstrated that in the majority of
of feedback training sessions the separability between motor
imagery-related brain states can be reinforced and a good
BCI performance can be expected. In tetraplegic patients the
situation is less clear. Only in one patient motor imagery-
related EEG patterns could be discriminated in the initial
training session. Here extensive trainings sessions without
and with feedback are necessary to achieve a satisfied BCI
performance at least in some patients.
This work was supported by Wings for Life—Spinal Cord
Research Foundation (002/06), EU Project Presenccia 27731,
EU COST BM0601 Neuromath, “Allgemeine Unfallver-
sicherung AUVA”, and Lorenz B¨ ohler Gesellschaft.
 E.Gerardin, A.Sirigu,S. Leh´ ericy, etal., “Partially overlapping
neural networks for real and imagined hand movements,”
Cerebral Cortex, vol. 10, no. 11, pp. 1093–1104, 2000.
 H. H. Ehrsson, J. P. Kuhtz-Buschbeck, and H. Forssberg,
“Brain regions controlling nonsynergistic versus synergistic
movement of the digits: a functional magnetic resonance
imaging study,” Journal of Neuroscience, vol. 22, no. 12, pp.
 G. Pfurtscheller and F. H. Lopes da Silva, “Event-related
EEG/MEG synchronization and desynchronization: basic
principles,” Clinical Neurophysiology, vol. 110, no. 11, pp.
 J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller,
and T. M. Vaughan, “Brain-computer interfaces for commu-
nication and control,” Clinical Neurophysiology, vol. 113, no.
6, pp. 767–791, 2002.
 G. Pfurtscheller, G. R. M¨ uller-Putz, R. Scherer, and C. Neuper,
“Rehabilitation with brain-computer interface systems,” Com-
puter, vol. 41, no. 10, pp. 58–65, 2008.
 R. Salmelin and R. Hari, “Spatiotemporal characteristics
of sensorimotor neuromagnetic rhythms related to thumb
movement,” Neuroscience, vol. 60, no. 2, pp. 537–550, 1994.
 L. Leocani, C. Toro, P. Manganotti, P. Zhuang, and M. Hallett,
“Event-related coherence and event-related desynchroniza-
tion/synchronization in the 10Hz and 20Hz EEG during
self-paced movements,” Electroencephalography and Clinical
Neurophysiology/Evoked Potentials Section, vol. 104, no. 3, pp.
 P. Derambure, L. Defebvre, J. L. Bourriez, F. Cassim, and J.
D. Guieu, “Event-related desynchronization and synchroniza-
tion. Reactivity of cortical electroencephalographic rhythms
Clinical Neurophysiology, vol. 29, no. 1, pp. 53–70, 1999.
 G. Pfurtscheller and C. Neuper, “Motor imagery activates
primary sensorimotor area in humans,” Neuroscience Letters,
vol. 239, no. 2-3, pp. 65–68, 1997.
 C. Neuper and G. Pfurtscheller, “Motor imagery and ERD,”
in Event-Related Desynchronization, G. Pfurtscheller and F. H.
Lopes da Silva, Eds., vol. 6 of Handbook of Electroencephalog-
raphy & Clinical Neurophysiology, pp. 303–325, Elsevier,
Amsterdam, The Netherlands, 1999.
 H.Ramoser,J.M¨ uller-Gerking,andG.Pfurtscheller,“Optimal
spatial filtering of single trial EEG during imagined hand
movement,” IEEE Transactions on Rehabilitation Engineering,
vol. 8, no. 4, pp. 441–446, 2000.
 G.Pfurtscheller,R.Scherer,G.R.M¨ uller-Putz,andF.H.Lopes
da Silva, “Short-lived brain state after cued motor imagery in
naive subjects,” European Journal of Neuroscience, vol. 28, no.
7, pp. 1419–1426, 2008.
 G. R. M¨ uller-Putz, R. Scherer, C. Brunner, R. Leeb, and G.
Pfurtscheller, “Better than random? A closer look on BCI
results,” International Journal of Bioelectromagnetism, vol. 10,
no. 1, pp. 52–55, 2008.
 J. M¨ uller-Gerking, G. Pfurtscheller, and H. Flyvbjerg, “Clas-
sification of movement-related EEG in a memorized delay
task experiment,” Clinical Neurophysiology, vol. 111, no. 8, pp.
 B. Hjorth, “An on line transformation of EEG scalp potentials
into orthogonal source derivations,” Electroencephalography
 J. Perelmouter and N. Birbaumer, “A binary spelling interface
with random errors,” IEEE Transactions on Rehabilitation
Engineering, vol. 8, no. 2, pp. 227–232, 2000.
 B. Blankertz, G. Dornhege, M. Krauledat, K.-R. M¨ uller, and
fast acquisition of effective performance in untrained sub-
jects,” NeuroImage, vol. 37, no. 2, pp. 539–550, 2007.
 G. Pfurtscheller, C. Neuper, and N. Birbaumer, “Human
brain-computer interface,” in Motor Cortex in Voluntary
Movements, A. Riehle and E. Vaadia, Eds., pp. 367–401, CRC
Press, Boca Raton, Fla, USA, 2005.
 C. Neuper, R. Scherer, M. Reiner, and G. Pfurtscheller,
“Imagery of motor actions: differential effects of kinesthetic
and visual-motor mode of imagery in single-trial EEG,”
Cognitive Brain Research, vol. 25, no. 3, pp. 668–677, 2005.
 G. Pfurtscheller, R. Leeb, C. Keinrath, et al., “Walking from
thought,” Brain Research, vol. 1071, no. 1, pp. 145–152, 2006.
 H. Alkadhi, P. Brugger, S. H. Boendermaker, et al., “What
disconnection tells about motor imagery: evidence from
paraplegic patients,” Cerebral Cortex, vol. 15, no. 2, pp. 131–
 P. Sabbah, S. de Schonen, C. Leveque, et al., “Sensorimotor
cortical activity in patients with complete spinal cord injury:
a functional magnetic resonance imaging study,” Journal of
Neurotrauma, vol. 19, no. 1, pp. 53–60, 2002.
 G. Pfurtscheller, C. Brunner, A. Schl¨ ogl, and F. H. Lopes da
Silva, “Mu rhythm (de)synchronization and EEG single-trial
classification of different motor imagery tasks,” NeuroImage,
vol. 31, no. 1, pp. 153–159, 2006.