Preliminary Study to Detect Gait Initiation Intention Through a BCI

System

Daniel Planelles, Enrique Hortal, Eduardo I

anez,

Alvaro Costa, Andr

Ubeda and Jos

e M. Azor

ın

Biomedical Neuroengineering Group, Miguel Hern

andez University of Elche, Avda. Universidad, Elche (Alicante), Spain

Keywords:

Brain-Computer Interface, Movement Intention, Gait.

Abstract:

In this paper is presented an experiment designed to detect the will to perform several steps forward (as

walking onset) before it occurs using the electroencephalographic (EEG) signals collected from the scalp. The

preliminary results from ﬁve users have been presented. In order to improve the quality of the signals acquired

some different spatial ﬁlters are applied and compared. In the future, the improved Brain-Computer Interface

of this paper will be used as part of the control system of an exoskeleton attached to the lower limb of people

with incomplete and complete spinal cord injury to initiate their gait cycle.

1 INTRODUCTION

There is a range of longer-term problems that a person

who has suffered a stroke might continue to face after

they have left hospital. Patients with stroke normally

have communication, cognitive, emotional, physical

or visual problems so they need attendance or even

rehabilitation. However, there is other disabilities as

people with spinal cord injury or multiple sclerosis

that involve lost of voluntary mobility. Therefore, the

research community has to do a huge effort to ﬁnd so-

lutions to restore their capacities or, at least, to facil-

itate new technology with the aim of improving their

lives.

For this reason, Brain-Computer Interfaces (BCIs)

have seen a rapid development during the last years as

an assistive technology. BCIs are an alternative com-

munication method for people with a severe motor

disability as they allow generating control commands

with the only help of the thoughts (Guido Dornhege

and Muller, 2007; Nicolelis, 2001). Therefore, BCIs

could increase their independence and also could im-

prove their quality of life.

Current technology allows collecting and process-

ing EEG signals that occur just before performing an

action and thus we can know the intention to perform

a movement (Bai et al., 2011). It can be used to assist

subject movements whenever he/she wishes which is

one of the great practical advantages of this approach.

In a motor rehabilitation process, it can make a big

improvement since it would be possible to share ef-

forts between the subject and, for example, an ex-

oskeleton attached to the lower limb (Moreno et al.,

2011). This coordination between the will to exe-

cute a movement and the performance of the action

itself increases the likelihood of the brain to create

new communication channels due to neuronal plas-

ticity (Kolb et al., 2011). Through this, the effects

of rehabilitation increase a greater extent in a much

shorter time frame.

This paper has been developed under the BioMot

project - Smart Wearable Robots with Bioinspired

Sensory-Motor Skills (FP7-ICT-2013-10) funded by

the Commission of the European Union, which pre-

tends to control an exoskeleton attached to the lower

limb of the disabled user capturing and processing

their electroencephalographic (EEG) signals. As part

of this challenge, our purpose is to activate the con-

trolled gait cycle of the exoskeleton using the inten-

tion to walk of the user. Therefore, the EEG signals

will be analysed in order to detect the intention of the

gait onset and this will be turn into an activation com-

mand of the exoskeletons engines or a stimulus over

the leg muscles with Functional Electrical Stimula-

tion (FES). At the moment, the focus is on the system

to detect the intention to start walking and initially,

only healthy subjects have performed the test.

For our knowledge, there are two phenomena ex-

tensively used in BCI related to the motor inten-

tion. On the one hand, a kind of Movement-Related

Cortical Potential (MRCP) which is a slow poten-

tial called Bereitschaftspotential or readiness poten-

tial (Shibasaki and Hallett, 2006) but this paper will

not take into consideration this phenomenon. On

Planelles D., Hortal E., Iánez E., Costa Á., Úbeda A. and M. Azorín J..

Preliminary Study to Detect Gait Initiation Intention Through a BCI System.

DOI: 10.5220/0005167800610066

In Proceedings of the 2nd International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX-2014), pages 61-66

ISBN: 978-989-758-056-7

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

the other hand, the event-related desynchronization

(ERD) characterized by a decrease in the spectral

power of EEG signals in mu and beta frequency bands

(Pfurtscheller and da Silva, 1999). This phenomenon

starts up to 2 seconds before the movement onset and

it ends more or less when the movement is ﬁnished.

After that, the spectral power recovers its magnitude

generating the event-related synchronization (ERS).

2 MATERIALS AND METHODS

2.1 Test Description

Five healthy subjects between 22 and 29 years old

(26.50±3.15) performed one session. All voluntaries

had normal vision and hearing and no history of neu-

rological or psychiatric disorders. Each subject was

instructed to remain at rest on their two legs during

at least 5 seconds to have enough information to be

used as resting time (or baseline) and then to perform

several steps forward. However, the user could begin

the movement whenever they want after the restricted

period (limited to 10 seconds). No interface guides

the subject but they know when a task starts through

an advice from the experimenter.

8 runs with 10 repetitions each have been per-

formed. Between each run there is one or two min-

utes of break time and between repetitions there is a

few seconds. The user has to wear several measure-

ment equipments to register the EEG signals and the

kinematic of the user. Then, a cap with EEG elec-

trodes, a backpack with the electronic devices that

registers the EEG signals and 7 Inertial Measurement

Units (IMUs) distributed over the lower limb are used.

These devices are connected to a laptop which is over

a cart managed by the experimenter. Therefore, the

experimenter pulls the cart while the user is moving

forward to keep around 1.5 meters between the user

and the cart. In Fig. 1 it is possible to see an example

of test.

2.2 Experimental Set Up

The system architecture is composed by a Brain-

Computer Interface (BCI) that will capture and pro-

cess EEG signals to command, in the future, an ex-

oskeleton attached to the lower limb and seven Iner-

tial Measurement Units (IMUs) managed by the Mo-

tion Capture System (Technaid S.L.), which are dis-

tributed over the lower body to register kinematics.

Both acquisition systems are synchronized.

Figure 1: The user is wearing the cap, the backpack and the

Motion Capture System while the experimenter follows him

with the cart and the computer.

2.2.1 Brain-Computer Interface

The EEG signals are registered through 32 active

electrodes. These electrodes are placed using a cap

(g.GAMMAcap, g.tec medical engineering, GmbH,

Austria). This cap is very useful as it allows an easy

placement of the electrodes. The 32 electrodes are

placed following a uniform distribution over the scalp.

According to the International 10-10 System, the po-

sition of the electrodes is the following: FZ, FC5,

FC3, FC1, FCZ, FC2, FC4, FC6, C5, C3, C1, CZ,

C2, C4, C6, CP5, CP3, CP1, CPZ, CP2, CP4, CP6,

P3, P1, PZ, P2, P4, PO7, PO3, POZ, PO4 and PO8.

This distribution of the electrodes is shown in Fig. 2.

These electrodes are the g.LADYbird model, sintered

Ag/AgCl crown with a 2-pin safety connector. These

electrodes need a conductive gel that comes in contact

the scalp with the sensor. The ground sensor is located

in AFz and the reference is placed on the earlobe. The

signals of the 32 electrodes are acquired through two

commercial g.USBamp devices from g.tec synchro-

nized by using g.INTERsync device and they have

two preampliﬁers g.GAMMAbox. Each ampliﬁer has

16 channels and the sampling frequency used to reg-

ister the signals is 1200 Hz.

NEUROTECHNIX2014-InternationalCongressonNeurotechnology,ElectronicsandInformatics

Figure 2: Placement of the 32 electrodes over frontal and

parietal lobes.

2.2.2 Motion Capture System

The motion capture system Tech MCS is a com-

plete wireless motion analysis system. It manages

the seven IMUs of the company Technaid which are

used in our experiments and they are placed as Fig. 3

shows. The sampling frequency used is 30 Hz. Each

Tech IMU integrates three different types of sensors

as an accelerometer, a gyroscope and a magnetometer.

A sophisticated and robust algorithm, calibrated also

taking into account changes in temperature, results in

a very precise and robust estimation of 3D orientation,

even during changing environmental conditions.

2.3 Signal Processing

The EEG signals voltage is around a few microvolts,

consequently the signals are easily affected by other

sources of voltage that are not the cerebral activity as

ocular or muscular movements. Therefore, it is nec-

essary to reduce the undesirable contribution of each

electrode using some temporal and frequency ﬁlters.

In this sense, the EEG data is ﬁltered with two fre-

quency ﬁlters that eliminate the power line interfer-

ence, the DC component and some artifacts. Then, a

50 Hz Notch ﬁlter and a 4th order Butterworth from

1 to 100 Hz ﬁlter are used. Thus, the information of

mu and beta (8-30 Hz) and surrounding frequencies

are isolated. Moreover, due to the proximity of the

EEG electrodes and the numerous neural connections,

the signal acquired per each electrode is partially af-

fected by the potential produced in other location of

the scalp. Hence, in this paper two different spatial

ﬁlters are used to reduce that neighbour contribution.

On the one hand, a Common Average Reference

Figure 3: IMUs distribution. 100: Tech HUB; 122: Lum-

bar; 123: Right quadriceps; 124: Right biceps; 125: Right

foot; 130: Left quadriceps; 133: Left biceps; 134: Left foot.

(CAR) is used (Alhaddad, 2012). This spatial ﬁlter

consists of subtract the 32 EEG channels mean data

to each channel. Therefore, it is expected that the in-

formation processed per electrode comes from itself,

reducing the contribution of the remaining electrodes.

On the other hand, a Laplacian (LAP) algorithm

is applied for all electrodes (McFarland et al., 1997).

This algorithm uses the information received from

all the remaining electrodes and their distances from

them. The visual result is a smoother time signal

which should contain only the contribution coming

from the particular position of the electrode. The

Laplacian is computed according to the formula:

LAP

= Vi

−

∑

jεSi

i j

V j

(1)

where Vi

LAP

is the result of applying this algorithm

to the electrode i, Vi

is the electrode i signal before

the transformation and,

i j

∑

jεSi

i j

(2)

where Si contains all the electrodes except from the

electrode i and d

i j

is the distance between i and j elec-

trodes.

In order to compare, a third option without any

spatial ﬁlter (NOF) is used.

2.4 Features Extraction and Classiﬁer

The EEG data are cut from 5 seconds before each

movement onset sample to 1 second after it. Only

PreliminaryStudytoDetectGaitInitiationIntentionThroughaBCISystem

these data will be analysed to distinguish between

premovement and resting time. Then, the ﬁrst three

seconds will be used as resting data and the last three

seconds as premovement data. To know the starting

sample, the data acquired from the IMUs are used. An

automatic searching method based on the ﬁrst change

in the curve obtained from IMUs is implemented. The

number of IMU is higher than needed in this experi-

ment because other lines of research in BioMot take

under consideration kinematics.

Later, the selected EEG data per each task is fed

into a 7th order autoregressive process that returns a

vector with the estimated coefﬁcients a of a linear re-

gression model per a EEG channel. This vector is

calculated as follows:

x(n) = 1 +

∑

i=1

x(n −i) (3)

where x is the EEG data and a is the vector of

estimated coefﬁcients.

Therefore, these coefﬁcients are used to train a

Support Vector Machine (SVM) classiﬁer. SVM is an

approach where the objective is to ﬁnd the best sepa-

ration hyperplane, which provides the highest margin

distance between the nearest points of the two classes

to separate them.

3 RESULTS

The system is tested using an 8-fold cross validation

with each run as a fold. This statistic analysis per-

forms all combinations of 7 runs to train the SVM

classiﬁer and uses the other one to test it. In each iter-

ation, the number of correct and incorrect detections

are added and, at the end, the following three statisti-

cal index are calculated:

T PR =

Premovement data correctly detected

Number of premovement data

∗ 100

(4)

FPR =

Resting data incorrectly detected

Number of resting data

∗ 100 (5)

ACC =

Tasks correctly detected

Number of tasks performed

∗ 100 (6)

The results obtained per each user with the two

spatial ﬁlters and no one are summarized in Table

1. In view of the results, it is possible to say that

when a spatial ﬁlter is used the FPR and the ACC are

Figure 4: False Positive Rate (FPR) in all users with the

different ﬁlters: no ﬁlter (NOF), laplacian ﬁlter (LAP) and

common average reference ﬁlter (CAR).

Figure 5: Accuracy Rate (ACC) in all users with the differ-

ent ﬁlters: no ﬁlter (NOF), laplacian ﬁlter (LAP) and com-

mon average reference ﬁlter (CAR).

improved (around 11% and 6% respectively), except

user E who is an abnormal case as it is shown in ﬁg-

ures 4 and 5. However, there is not signiﬁcant differ-

ences in TPR average. Leaving out user E and NOF,

TPR is better using LAP ﬁlter but FPR is lower us-

ing CAR and at the end, the accuracy of the system

is similar with both ﬁlters. Therefore, the difference

between LAP and CAR ﬁlter is not signiﬁcant but at

least, according to ACC, it is better than NOF. In aver-

age, TPR and FPR are around 68% and 34% (respec-

tively) using any ﬁlter. However, there is huge differ-

ences between users and for example, user B reaches

93.75% of TPR while user D only achieves 51.56%

with the same spatial ﬁlter. Then, the system has to

be improved in both index with more complex signal

processing to achieve better features and a more sta-

ble system.

NEUROTECHNIX2014-InternationalCongressonNeurotechnology,ElectronicsandInformatics

Table 1: Results of the 8-fold cross validation.

TRUE FALSE ACCURACY

POSITIVE RATE POSITIVE RATE RATE

USER NOF LAP CAR NOF LAP CAR NOF LAP CAR

A 62.50 69.44 66.67 41.67 22.22 19.44 60.42 73.61 73.61

B 77.08 89.58 93.75 60.08 43.75 43.75 58.33 72.92 75.00

C 80.56 77.78 66.67 44.44 26.39 26.39 68.06 75.69 70.14

D 70.31 59.38 51.56 48.44 26.56 21.88 60.94 66.41 64.84

E 48.21 46.43 58.93 30.36 53.57 53.57 58.93 46.43 52.68

MEAN 67.73 68.52 67.52 45.00 34.50 33.01 61.34 67.01 67.25

4 CONCLUSIONS AND FUTURE

WORKS

In this paper a system to detect the intention to start

walking has been presented. According to the results

obtained from ﬁve healthy users, the system has to

be improved to achieve better TPR and FPR. If the

system is able to predict each movement with a re-

ally low FP rate, the classiﬁer output could serve as a

command to activate the engines of an exoskeleton or

a FES system to start walking. Therefore, in a rehabil-

itation process an exoskeleton could be used to sup-

port the lower limbs while the user carries out men-

tally walking intentions. The relationship between the

cognitive process to perform such movement and the

real movement could improve the rehabilitation due

to cerebral plasticity.

Therefore, other methods to characterize the EEG

signals before the movement onset will be studied.

For example, it is possible to calculate the power

spectral and to extract the best component of each

user to be used as feature. Moreover, a frequency ﬁl-

ter narrower could be applied in order to isolate the

mu and beta frequency bands which are involved in

ERD phenomenon. Furthermore, ERD can be used

purely as theory indicates (a relative decrement of

power spectral in a special frequency band measured

in percentage). Usually, EEG signals have some arti-

facts like eye blinks or muscle activity that should be

removed. In particular, eye movements could be re-

moved from EEG signals recording the electroculag-

raphy (EOG) signals using 2 bipolar electrodes (hor-

izontal and vertical channels) and then reducing their

contribution through a linear regression which re-

lates EEG and EOG signals (Kenemans et al., 1991).

Moreover, artifacts due to neck movements can be

easily seen in EEG signals due to their magnitude, so

it is possible to add a voltage threshold to avoid it.

The population of this experiment and also the

number of sessions will be increased. Furthermore,

patients with complete and incomplete spinal cord in-

jury will perform the experiment in order to test the

system and to evaluate the performance. It is ex-

pected that some patients keep their brain procedures

related to the intention of movement as healthy people

although they could be weaker or allocated in other

brain area (Wei et al., 2011). Then, a real-time test

with a better system will be performed both healthy

as patient users.

ACKNOWLEDGEMENTS

This research has been funded by the Commission of

the European Union under the BioMot project - Smart

Wearable Robots with Bioinspired Sensory-Motor

Skills (Grant Agreement number IFP7-ICT- 2013-10-

611695), by the Spanish Ministry of Economy and

Competitiveness as part of the Brain2motion project

- Development of a Multimodal Brain-Neural Inter-

face to Control an Exoskeletal: Neuroprosthesis Hy-

brid Robotic System for the Upper Limb (DPI2011-

27022-C02-01), and by Conselleria d’Educaci

o, Cul-

tura i Esport of Generalitat Valenciana of Spain

through grant VALi+d ACIF/2012/135.

REFERENCES

Alhaddad, M. J. (2012). Common Average Reference

(CAR) Improves P300 Speller. International Journal

of Engineering and Technology, 2.

Bai, O., Rathi, V., Lin, P., Huang, D., Battapady, H., Fei,

D.-Y., Schneider, L., Houdayer, E., Chen, X., and

Hallett, M. (2011). Prediction of human voluntary

movement before it occurs. Clinical Neurophysiology,

122(2):364 – 372.

Guido Dornhege, Jos

e del R. Mill

an, T. H. D. J. M. and

Muller, K.-R. (2007). Towards Brain-Computer Inter-

facing. The MIT Press.

Kenemans, J. L., Molenaar, P. C., Verbaten, M. N., and

Slangen, J. L. (1991). Removal of the ocular artifact

PreliminaryStudytoDetectGaitInitiationIntentionThroughaBCISystem

from the eeg: A comparison of time and frequency

domain methods with simulated and real data. Psy-

chophysiology, 28:114–121.

Kolb, B., Muhammad, A., and Gibb, R. (2011). Searching

for factors underlying cerebral plasticity in the normal

and injured brain. Journal of Communication Disor-

ders, 44:503–514.

McFarland, D. J., McCane, L. M., David, S. V., and Wol-

paw, J. R. (1997). Spatial ﬁlter selection for EEG-

based communication. Electroencephalography and

Clinical Neurophysiology, 103:386–394.

Moreno, J. C., Ama, A. J., Reyes-Guzmn, A., Gil-Agudo,

n., Ceres, R., and Pons, J. L. (2011). Neurorobotic

and hybrid management of lower limb motor disor-

ders: a review. Medical and Biological Engineering

and Computing, 49(10):1119–1130.

Nicolelis, M. A. L. (2001). Actions from thoughts. Nature,

(409):403407.

Pfurtscheller, G. and da Silva, F. H. L. (1999). Event-

related EEG/MEG synchronization and desynchro-

nization: basic principles. Clinical Neurophysiology,

110:1842–1857.

Shibasaki, H. and Hallett, M. (2006). What is the

Bereitschaftspotential? Clinical Neurophysiology,

117:2341–2356.

Wei, L., Yue, H., Jiang, X., and He, J. (2011). Brain ac-

tivity during walking in patient with spinal cord in-

jury. International Symposium on Bioelectronics and

Bioinformatics, pages 96–99.

NEUROTECHNIX2014-InternationalCongressonNeurotechnology,ElectronicsandInformatics