Content uploaded by Jaime Ibáñez
Author content
All content in this area was uploaded by Jaime Ibáñez on May 17, 2015
Content may be subject to copyright.
Upper-limb muscular electrical stimulation driven by EEG-based detections
of the intentions to move: A proposed intervention for patients with stroke
J Ib´
a˜
nez1, JI Serrano1, MD del Castillo1, E Monge2, F Molina2,
FM Rivas2, I Alguacil2, JC Miangolarra2and JL Pons1
Abstract— This study proposes an intervention for stroke
patients in which electrical stimulation of muscles in the affected
arm is supplied when movement intention is detected from
the electroencephalographic signal. The detection relies on the
combined analysis of two movement related cortical patterns:
the event-related desynchronization and the bereitschaftspo-
tential. Results with two healthy subjects and three chronic
stroke patients show that reliable EEG-based estimations of
the movement onsets can be generated (on average, 66.9 ±
26.4 % of the movements are detected with 0.42 ±0.17 false
activations per minute) which in turn give rise to electrical
stimuli providing sensory feedback tightly associated to the
movement planning (average detection latency of the onsets
of the movements was 54.4 ±287.9 ms).
I. INTRODUCTION
The damage of neural networks in the brain caused by
stroke may affect the functionality of limbs on one side
of the body. Successfully recovering the ability to perform
functional tasks with the affected limb depends mainly on
the characteristics of the brain injury (size and location),
and on the effectiveness of the rehabilitation therapy [1].
Around 30% of chronic stroke patients do not recover the
arm and hand functionality despite intensive treatment and
rehabilitation [2].
Novel therapies focusing on the neural rehabilitation of the
patients may lead to an improvement of their condition in the
long term. The use of Brain-computer interface (BCI) tech-
nology based on the electroencephalographic (EEG) activity
has gained interest in this regard during the last years [3]. The
EEG activity allows characterizing movement-related mental
states at the exact moment they occur in the brain [4], which
in turn may be used to supply patients with sensory feedback
coupled with their expectations of movement. Supplying
such tightly associated in time sensory feedback is postulated
to induce plasticity in cortical regions targeting the damaged
limb, which in turn leads to restoration of normal motor
control [3]. In this regard, recent studies have proven the
relevance of the proprioceptive feedback timing to achieve
associative neural facilitation [5].
The Bereitschaftspotential (BP) is defined as a slow neg-
ative shift of the EEG amplitude over the central cortical
areas that precedes voluntary movements (for a review
see [6]). The BP has been used in previous studies to
1J. Ib´
a˜
nez, M. D. del Castillo, J. I. Serrano and J. L. Pons are with the
Bioengineering Group of the Spanish Research Council (CSIC), 28500 La
Poveda, Arganda del Rey, Spain jaime.ibanez at csic.es
2E. Monge, F. Molina, F. M. Rivas, I. Alguacil and J. C. Miangolarra are
with the LAMBECOM group of the Universidad Rey Juan Carlos, Alcorc´
on,
Spain
detect online the onsets of self-paced ankle dorsiflexions
[4], and to trigger peripheral nerve stimulation during self-
paced imagined movements [5]. In that case, it was showed
that increased cortical excitability could be induced with a
correct function of the EEG-based system. Although positive
results have been achieved in the use of the BP to detect
onsets of voluntary movements in healthy subjects, some
limitations may also be considered. Firstly, the amplitude
of the BP (5-10 µV) and the frequency band where it is
contained (0.05-1 Hz) make this cortical pattern vulnerable
to noisy environments and artifacts. Secondly, in patients
with cortical damage such as stroke patients, altered BP
patterns may be observed [7], which may affect the single-
trial detection of the BP as compared to healthy subjects [8].
A possible way of boosting EEG-based systems aimed to
detect the intention to move is to combine the BP with other
EEG movement-related patterns providing complementary
information. The event-related desynchronization (ERD) is
a well-documented movement-related cortical pattern. For
voluntary movements performed with the upper-limb, the
ERD consists in a decrease of EEG signal power in the
contralateral alpha (8-12 Hz) and beta (13-30 Hz) rhythms
starting around 2 s before the onset of voluntary movements
[9]. As in the case of the BP, the spatio-temporo-frequential
distribution of the ERD observed when averaging a number
of EEG segments preceding voluntary movements shows a
distinguishable pattern [10], which may be useful to locate
the onsets of these movements. In fact, previous studies
have used the ERD pattern to anticipate voluntary movement
events [11]. As in the analysis of the BP, the ERD pattern
of stroke patients presents variations with respect to healthy
subjects [12]. Therefore, it is of interest to study how stroke-
related cortical changes may affect a BCI driven by these two
different cortical patterns.
Here it is proposed an intervention for stroke patients in
which electrical stimulation of upper-limb muscles is deliv-
ered according to EEG-based estimations of the onsets of
voluntary movements. To this end, the sensorimotor rhythms
and the slow cortical potentials are analysed. The combined
function of the BCI system and the electrical stimuli is tested
in single experimental sessions with healthy subjects and
chronic stroke patients. To the authors’ knowledge, this is
the first time that precise temporal estimations regarding
movement intentions are used to close the BCI loop with
electrical stimulation in upper-limb functional movements
performed by stroke patients. Results given are expected to
serve as a partial validation of such intervention.
II. MET HODS
A. Participants
Two healthy subjects (right handed, 28 and 33 years
old) and three chronic stroke patients were recruited to
validate the proposed system. The patients’ description can
be found in Table I. None of the subjects measured had prior
experience with BCI paradigms. The experimental protocol
was approved by the Ethical Committee of the ‘‘Universidad
Rey Juan Carlos’’ (Madrid), and warranted its accordance
with the Declaration of Helsinki. All patients signed a written
informed consent.
Code Age Gender Stroke Side F-M
Type
P01 69 M Hemorrg R 64
P02 52 F Isquemic L 126
P03 54 M Isquemic L 68
TABLE I
DES CRIP TON OF T HE S TRO KE PATI EN TS ’CONDITIONS. (F-M RE FE RS T O
F¨
UGL-MEY ER S CA LE )
B. Apparatus and experimental protocol
Movements of the affected arm (dominant arm for healthy
subjects) were measured with three solid-state gyroscopes
(Technaid S.L., Madrid, Spain), placed on the hand dorsum,
the distal third of the forearm, and the middle of the arm.
The data were sampled at 100 Hz.
EEG signals were recorded from AFz, F3, F1, Fz, F2,
F4, FC3, FC1, FCz, FC2, FC4, C5, C3, C1, Cz, C2, C4,
C6, CP3, CP1, CPz, CP2, CP4, P3, P1, Pz, P2, P4, PO3,
PO4 and Oz, (according to the international 10-20 system)
using active Ag/AgCl electrodes (Acticap, Brain Products
GmbH, Germany). The reference was set to the voltage of
the earlobe contralateral to the arm moved. AFz was used
as ground. The signal was amplified (gUSBamp, g.Tecgmbh,
Austria) and sampled at 256 Hz.
The electrical stimuli were delivered at the anterior del-
toids and triceps with a multichannel monopolar neurostim-
ulator with charge compensated pulses (UNA Systems,
Belgrade, Serbia). The common electrode was located at
the oleocranon. Sub motor-threshold stimulation was used.
Pulse width and frequency were set to 250 µs and 30 pps,
respectively. The stimulator was controlled by a stand alone
computer (OS QNX Software Systems, Ottawa, Canada) that
received activation commands from the computer recording
the EEG activity via a digital signal.
Each participant was measured during one session. Partic-
ipants sat in a comfortable chair with their arms supported
on a table. They were instructed to remain relaxed with their
eyes open and their gaze fixated on a point on the wall.
In the first part of the session (“TRAIN”), the participants
were asked to perform self-paced reaching movements with
the affected arm (the dominant arm for control subjects).
During the resting periods between movements, participants
were asked to remain relaxed for around 5-8 s. Participants
performed 30 movements in this part. The data recorded was
used to train an EEG-based detector of the onsets reaching
movements. In the second part of the session (“INTERV”)
participants were asked to stare at a screen that presented
three states in each trial. First the word “Rest” was printed
on the screen for a variable period of time until the EEG-
based detector showed negative (non-movement) estimations
during at least 2 s. When that condition was reached, a blank
screen was showed indicating the participants that they could
initiate a movement whenever they wanted (trying to wait
more than 3 s). When the participants initiated a movement,
gyroscopic sensors detected it and the screen printed the
word “Movement” until the movement ended. At this point,
the trial was finished. During this process, electrical stimuli
were delivered each time the EEG-based detector estimated
that movement intention was detected, unless the “Rest”
state was present (during this state, detections of movement
intention were not listened). Participants were told that the
electrical stimuli appeared whenever motor-related mental
processes were observed. Electrical stimuli lasted 2 s if no
movement was performed (wrong detections) and they lasted
until the end of the movement if the movement was given
concurrently with the detection. A total of 60 movements
were performed during this phase.
C. Detection of the onset of the movements
To detect the actual onsets of the reaching movements, the
data from the gyroscopic sensor in the arm were analyzed.
Data were low-pass filtered (Butterworth, order 2, <6 Hz).
The threshold amplitude for the detection of the onsets of
the movements was set to 5 % of the peak amplitude in the
training data.
D. Description of the classifier architecture and validation
Two classifiers (based on the movement-related ERD and
BP patterns) were combined to estimate the instant at which
the onsets of the movements were located (see Fig. 1).
A na¨
ıve Bayes classifier was used to detect the ERD
pattern preceding the movements in each participant. Band-
pass filtering (Butterworth, 3th order, 0.5 Hz <f1, 35 >f2)
and small laplacian filter were used. The power values were
estimated in segments of 1.5 s and for frequencies between
7-30 Hz in steps of 1 Hz. Welch’s method was used to this
end (Hamming windows of 1 s, 50 % overlapping). Power es-
timations were generated every 100 ms. The values obtained
in the training run from -3 s to -0.5 s (with respect to the
movement onsets) were labelled as resting state examples,
whereas the estimations generated at t = 0 s were labelled as
movement onset examples. The Bhattacharyya distance was
used to select the 10 best features (channel/frequency pairs)
to build the Bayesian classifier. The Bayesian classifier was
trained with the training examples of the selected features
and it was applied to the test data generating estimations of
movement intention every 100 ms.
A similar procedure to the one proposed in [13] was used
to detect the BP. In this case, a finite impulse response
bandpass filter (0.05 Hz <f1, 1 Hz >f2, 15th order) was
used. Three virtual channels were obtained by subtracting
the average potential of channels F3, Fz, F4, C3, C4, P3,
Pz and P4 to channels C1, Cz and C2. The average BP was
computed for the three resulting channels using the training
data. The channel showing the highest absolute BP peak
was selected for the online BP-based detection of movement
onsets. Finally, a matched filter was designed using the pre-
viously selected channel. To this end, the average BP pattern
was extracted from -1.5 s to 0 s from the trials in the training
dataset. During the online function, the matched filter was
applied to the virtual channel of the validation dataset. The
BP-based online estimations of movement intention were
also made every 100 ms.
Outputs from ERD- and BP-based detectors were com-
bined using a logistic regression classifier. Training examples
of the resting condition were taken from estimations of the
two detectors between -3 s and -0.5 s with respect to the
movement onset. The output estimations of the ERD and the
BP classifiers at the movement onset were used to model the
movement state. The classifier generated estimations of the
intention to move every 100 ms.
Finally, a threshold was applied to the output of the detec-
tor to decide at each moment whether movement intention
was detected. The threshold was optimally obtained using
the training dataset, following the criterion of maximizing
the percentage of trials with a true positive (TP) and with no
false positives (FP). The TP was defined as the percentage
of trials with a movement detection contained in the time
interval from -0.75 s to +0.75 s with respect to the actual
onset estimated by the gyroscopes. Detections of the clas-
sifier during the resting intervals between movements were
considered FP (activations during the intervals in which the
screen showed the word “Rest” were also accounted as FP).
III. RES ULT S
The features selected by the ERD and BP classifiers from
the training dataset are presented in Table II. In the case
of the ERD-classifier, features from contralateral central
positions were selected to detect the movement onsets in the
control subjects, whereas for the patients, different patterns
of selected features were observed. The average BP curve
obtained from the training dataset showed that, in participants
C01 and C02, the BP peak was located at t =-1.0 ms and
t=123.1 ms, respectively, and in the patients, the BP peak
was located at t =396.6 ms (P01), t =119.6 ms (P02), and
t=56.4 ms (P03).
Online results of the proposed EEG-based detector are
shown in Table III. On average, 66.9 ±26.4 % of the
movements were correctly detected, giving rise to electrical
stimuli associated to the movement intention. The number
of false activations generated per minute (FP/min) was on
average 0.42 ±0.17, and the movement onsets were detected
with average latencies of 54.4 ±287.9 ms with respect to the
actual onsets of the movements located with the gyroscopic
sensors. Fig. 2 shows the latencies of all true positives with
respect to the actual onsets of the movements in each subject,
reflecting an organisation of the detections around the actual
movement onset in the two healthy subjects and in P01,
Small Laplacian
lter
PSD
estimation
Naïve Bayes
classier
Large Laplacian
lter
Matched
lter
Logistic
regression
Threshold
applied
EEG
activity
Peripheral
electrical
stimulation
Bandpass lter
fc1>0.5Hz
fc2<35Hz
Bandpass lter
fc1>0.05Hz
fc2<1Hz
Bereitshafts-
potential
Event-related
desynchronization
Fig. 1. Schematic representation of the system architecture
C01 C02 P01 P02 P3
C3/12Hz C3/12Hz CPz/22Hz C6/20Hz CPz/14Hz
CP2/16Hz C3/11Hz CPz/23Hz C6/32Hz CPz/15Hz
CP2/15Hz C3/21Hz C1/22Hz C6/19Hz CPz/13Hz
C3/23Hz C3/20Hz P3/12Hz C6/18Hz CPz/20Hz
CP2/17Hz C3/17Hz PO3/12Hz C4/18Hz CPz/16Hz
C2/14Hz C3/18Hz C1/23Hz C6/22Hz CP2/20Hz
FC1/14Hz C3/19Hz C1/21Hz C2/21Hz CP2/18Hz
C1/17Hz CP3/10Hz CPz/16Hz C4/19Hz Cz/23Hz
C1/16Hz CP3/12Hz CPz/15Hz P4/10Hz CP2/19Hz
FC1/20Hz C3/10Hz CPz/24Hz FC1/11Hz Cz/20Hz
C1 Cz Cz C2 Cz
TABLE II
FEATU RE S SE LE CT ED B Y TH E ER D (FI RS T 10 LINES)AND B P (L AS T
LI NE )CL AS SI FIE RS F OR E ACH PA RTI CI PANT ( C CONTROL, P PATIEN TS )
whereas delayed detections were obtained in most trials in
patients P02 and P03.
IV. DISCUSSION AND CONCLUSIONS
This article presents an intervention for the upper-limb of
stroke patients aimed at promoting associative facilitation at
the motor cortex. The study shows results of a paradigm in
which participants performed self-paced reaching movements
with the arm and proprioceptive feedback was delivered
by means of an electrical stimulator activated each time
movement intention states were found in the EEG signal.
At the same time, a visual feedback was presented to the
participants to guide them through the intervention. The
visual paradigm assured that the basal (resting) state was
Code TP (%) FP/min Latency (ms)
C01 60.0 0.29 60.6±267.0
C02 90.0 0.32 30.6±303.1
P01 93.2 0.49 -56.4±156.1
P02 28.1 0.30 25.0±233.3
P03 63.3 0.69 139.5±350.0
Average 66.9±26.4 0.42±0.17 54.4±287.9
TABLE III
RESULTS OBTAINED WITH ALL SUBJECTS AND AVERAGE RESULTS
C01 C02 P01 P02 P03
EEG-based detection
latency (s)
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
Fig. 2. Latencies of the true positives with respect to the actual onsets of
the movements
reached after each movement, although this did not affect to
the self-pace nature of the exercise. This way, a demanding
protocol was achieved, which required a constant concen-
tration of the participants in the task. This is the first time
that the EEG-based online detection of the onsets of upper-
limb movements in chronic stroke patients is used to trigger
electrical stimulation.
It is concluded that a successful detection of the initiation
of volitional actions with the arm can be obtained and that
the electrical stimulation help to improve the specificity of
the EEG-based detector (the number of false activations
generated was lower than that observed in similar studies,
e.g. [4]). Other authors have also provided evidence of
improved decoding of motor-related cortical activity when
the BCI sensorimotor feedback loop is closed [14]. Results
here sustain this hypothesis and extend it by providing
evidence of a maintained EEG-based reliable detection of
movement onsets with temporal accuracy when subjects re-
ceived electrical stimulation. Having a source of information
regarding patients’ intentions to move and with temporal
accuracy is expected to improve BCI-based therapies, since
it ensures the active participation of the patients in the
rehabilitation task and implies a close interaction between
movement-related cortical processes and external electrical
stimuli. Furthermore, all subjects, except for P02, reported
that they were able to control the electrical stimulation when
they imagined instead of performed the reaching movements,
which provides a rationale for using this kind of technology
in patients without movement capacity. Since the onset of
imagined movements cannot be precisely located, results
with imagined movements were not extracted and are there-
fore not provided.
Finally, differences were observed between the healthy
subjects and the patients. According to the features selected
in each case and to the location of the BP peak, these differ-
ences were associated with altered BP and ERD patterns in
the patients, which is in line with previous studies [7], [12].
In future studies, longitudinal experiments may be carried out
with the patients performing multiple intervention sessions
to study whether such approach results in improvements of
functional scales.
ACKNOWLEDGMENT
This work has been funded by grant from the Spanish
Ministry of Science and Innovation CONSOLIDER INGE-
NIO, project HYPER (CSD2009-00067), from Proyectos
Cero of FGCSIC, Obra Social la Caixa (CSIC), from Project
CP Walker (DPI2012-39133-C03-01) and from the project
PIE-201350E070.
REFERENCES
[1] T. H. Murphy and D. Corbett, “Plasticity during stroke recovery: from
synapse to behaviour.” Nature reviews. Neuroscience, vol. 10, no. 12,
pp. 861–72, Dec. 2009.
[2] S.-M. Lai, S. Studenski, P. W. Duncan, and S. Perera, “Persisting
Consequences of Stroke Measured by the Stroke Impact Scale,” Stroke,
vol. 33, no. 7, pp. 1840–1844, Jul. 2002.
[3] J. J. Daly and J. R. Wolpaw, “Brain-computer interfaces in neurolog-
ical rehabilitation,” The Lancet Neurology, vol. 7, no. 11, pp. 1032–
1043, Nov. 2008.
[4] R. Xu, N. Jiang, C. Lin, N. Mrachacz-Kersting, K. Dremstrup, and
D. Farina, “Enhanced low-latency detection of motor intention from
EEG for closed-loop brain-computer interface applications.” IEEE
transactions on bio-medical engineering, vol. 61, no. 2, pp. 288–96,
Feb. 2014.
[5] I. K. Niazi, N. Mrachacz-Kersting, N. Jiang, K. Dremstrup, and
D. Farina, “Peripheral electrical stimulation triggered by self-paced
detection of motor intention enhances motor evoked potentials.” IEEE
transactions on neural systems and rehabilitation engineering, vol. 20,
no. 4, pp. 595–604, Jul. 2012.
[6] H. Shibasaki and M. Hallett, “What is the Bereitschaftspotential?”
Clinical neurophysiology, vol. 117, no. 11, pp. 2341–56, Nov. 2006.
[7] J. J. Daly, Y. Fang, E. M. Perepezko, V. Siemionow, and G. H.
Yue, “Prolonged Cognitive Planning Time, Elevated Cognitive Effort ,
and Relationship to Coordination and Motor,” Rehabilitation, vol. 14,
no. 2, pp. 168–171, 2006.
[8] I. K. Niazi, N. Jiang, O. Tiberghien, J. r. F. k. Nielsen, K. Dremstrup,
and D. Farina, “Detection of movement intention from single-trial
movement-related cortical potentials.” Journal of neural engineering,
vol. 8, no. 6, p. 066009, Dec. 2011.
[9] G. Pfurtscheller and F. H. L. da Silva, “Event-related EEG/EMG
Synchronization and Desynchronization: Basic Principles,” Clinical
Neurophysiology, vol. 110, pp. 1842–1857, 1999.
[10] O. Bai, Z. Mari, S. Vorbach, and M. Hallett, “Asymmetric Spatiotem-
poral Patterns of Event-related Desynchronization preceding Voluntary
Sequential Finger Movements: a High-resolution EEG Study,” Clinical
Neurophysiology, vol. 116, pp. 1213–1221, 2005.
[11] J. Ib´
a˜
nez, J. Serrano, M. del Castillo, J. A. Gallego, and E. Rocon,
“Online detector of movement intention based on EEG - Application
in tremor patients,” Biomedical Signal Processing and Control, vol. 8,
no. 6, pp. 822–829, Nov. 2013.
[12] M. Stepien, J. Conradi, G. Waterstraat, F. U. Hohlefeld, G. Curio,
and V. V. Nikulin, “Event-related desynchronization of sensorimotor
EEG rhythms in hemiparetic patients with acute stroke.” Neuroscience
letters, Nov. 2010.
[13] M. Jochumsen, I. K. Niazi, N. Mrachacz-kersting, D. Farina, and
K. Dremstrup, “Detection and classification of movement-related
cortical potentials associated with task force and speed.” Journal of
neural engineering, vol. 10, no. 5, p. 056015, Aug. 2013.
[14] M. Gomez-Rodriguez, J. Peters, J. Hill, B. Scho lkopf, A. Gharabaghi,
and M. Grosse-Wentrup, “Closing the sensorimotor loop: Haptic
feedback facilitates decoding of arm movement imagery,” in Systems
Man and Cybernetics (SMC), 2010 IEEE International Conference on.
IEEE, 2010, pp. 121–126.
