Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon

Brain-Computer Interface Race

Eduardo G. Ponferrada

, Anastasia Sylaidi

and A. Aldo Faisal

1,2,3

Department of Bioengineering, Imperial College London, London, U.K.

Department of Computing, Imperial College London, London, U.K.

Data Science Institute, London, U.K.

Keywords:

Brain-Computer Interfaces, Machine Learning, Cybathlon.

Abstract:

Neuromotor diseases such as Amyotrophic Lateral Sclerosis or Multiple Sclerosis affect millions of people

throughout the globe by obstructing body movement and thereby any instrumental interaction with the world.

Brain Computer Interfaces (BCIs) hold the premise of re-routing signals around the damaged parts of the

nervous system to restore control. However, the ﬁeld still faces open challenges in training and practical

implementation for real-time usage which hampers its impact on patients. The Cybathlon Brain-Computer

Interface Race promotes the development of practical BCIs to facilitate clinical adoption. In this work we pre-

sent a competitive and data-efﬁcient BCI system to control the Cybathlon video game using motor imageries.

The platform achieves substantial performance while requiring a relatively small amount of training data, the-

reby accelerating the training phase. We employ a static band-pass ﬁlter and Common Spatial Patterns learnt

using supervised machine learning techniques to enable the discrimination between different motor imageries.

Log-variance features are extracted from the spatio-temporally ﬁltered EEG signals to ﬁt a Logistic Regression

classiﬁer, obtaining satisfying levels of decoding accuracy. The systems performance is evaluated online, on

the ﬁrst version of the Cybathlon Brain Runners game, controlling 3 commands with up to 60.03% accuracy

using a two-step hierarchical classiﬁer.

1 INTRODUCTION

Individuals suffering from severe neuromotor disor-

ders, such as Multiple Sclerosis, Spinal Cord In-

jury or Amyotrophic Lateral Sclerosis, face extra-

ordinary barriers in communicating with their envi-

ronment in their daily lives. Overcoming such bar-

riers through the use of technology, while achieving

successful adoption by the end-users remains an open

problem (Makin et al., 2017). Previous work in as-

sistive technology towards addressing this challenge

includes the development of a robotic arm control-

led by 3D eye-tracking which is able to assist rea-

ching movements (Maimon-Mor et al., 2017; Dzie-

mian et al., 2016; Tostado et al., 2016). This system

identiﬁes movement targets based on binocular gaze

points and selectively employs a trigger to activate the

robotic arm. In a follow-up study, different methods

to control such triggers were evaluated, inferring that

blink detection can achieve better performance in ro-

botic control than electromyography (EMG) or voice

(Noronha et al., 2017). Despite this available range of

triggering approaches, the overall robust control of ro-

botic interfaces, remains substantially difﬁcult for se-

verely paralyzed individuals, where movement is ex-

tremely limited. To address such cases, our lab has

previously developed an online non-invasive Brain-

Computer Interface (BCI) to control external envi-

ronmental entities with no muscle activity involved,

using deep learning techniques for multi-class closed-

loop motor imagery decoding (Ortega et al., 2018b;

Ortega et al., 2018a). Similar decoding strategies have

also been proven successful for ofﬂine BCIs in previ-

ous studies (Yang et al., 2015; Tabar and Halici, 2017;

Lu et al., 2017). Despite its decoding advantages,

this type of BCI uses supervised machine learning to

match brain activity to user intents, requiring a sig-

niﬁcant amount of training data obtained throughout

lengthy training sessions. However, long training ti-

mes prior to BCI control pose a rejection risk from the

end-users’ side and can impair a mainstream adoption

process.

To address this limitation, here, we develop a data-

efﬁcient BCI system, which maintains decoding per-

Ponferrada, E., Sylaidi, A. and Faisal, A.

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race.

DOI: 10.5220/0006962400210032

In Proceedings of the 6th International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX 2018), pages 21-32

ISBN: 978-989-758-326-1

formance, while reducing the amount of required trai-

ning data, thereby facilitating successful BCI adop-

tion in a clinical setting. Data-efﬁcient BCI sys-

tems have been previously proposed primarily for off-

line motor imagery decoding (Ferrante et al., 2015).

Generally, to this date, the majority of BCI rese-

arch is dominated by ofﬂine decoding studies (Yang

et al., 2015; Tabar and Halici, 2017; Lu et al., 2017;

Zhang L. and Wang, 2010; Obermaier et al., 2001),

with only a few online testing cases of non-invasive

BCIs (Holz et al., 2015). Our system builds upon

these emerging approaches.

The Cybathlon (Riener and Seward, 2014) is a

competition in which people with disabilities com-

pete in a series of motor control activities using ad-

vanced assistive technology. The Cybathlon’s Brain-

Computer Interface Race is one of the presented acti-

vity tests intended for individuals with neuromotor

disorders, that consists of a competitive racing video

game, during which participants have to control an

avatar using strictly their brain activity.

Our system is particularly focused towards deve-

loping a competitive platform for the Cybathlon BCI

Race task, motivated by encouraging recent online de-

coding results within the same context (Ortega et al.,

2018a) (Schwarz et al., 2016).

The data-efﬁcient BCI system we developed in

order to tackle the Cybathlon challenge, is based

on Common Spatial Patterns (Ramoser et al., 2000)

(CSP) and a logistic regression classiﬁer (Tomioka

et al., 2007) to decode motor imageries.

The video game mechanics are simple. The players

of the game run on a given track at steady velocity

throughout the whole duration of the race. The track

occasionally changes colors to indicate switches in

desired commands and the players have to execute

these commands efﬁciently in order to obtain an ad-

vantage. Executing the wrong command whilst on a

coloured section, as well as sending any command

whilst on the grey (baseline) area of the track will

make the avatar fall behind by a few seconds. The

ﬁrst player reaching the ﬁnish line wins the game.

Our system was developed and adjusted for the

ﬁrst version of the Cybathlon Brain Runners video-

game. It is worth noting that this version was the

ﬁrst release of the Cybathlon videogame, which be-

came signiﬁcantly harder in the ﬁnal version used in

the competition.

The Cybathlon BCI Race rules state that the game

must be controlled using strictly brain activity. We

decided to control the execution of each one of the

commands using motor imageries, namely, thoughts

of executing different intuitive body movements. We

thus developed a system able to decode motor thoug-

hts and execute their associated video game command

in real time. We treated the no-command -necessary

when the avatar is on the grey sections of the track-

as an additional command, amounting to four alter-

native motor imageries altogether, available for deco-

ding. The motor imageries we decoded for control

purposes consisted on imaginations of opening and

closing the right hand, opening and closing the left

hand, ﬂexing and relaxing both feet at the ankle (dor-

siﬂexion) and contracting and relaxing the abdominal

muscles. Figure 1 presents the association between

section colour and motor imageries.

Figure 1: Association of game events with motor imageries.

2 METHODS

The BCI system is composed of two main modules.

The ﬁrst one is the ofﬂine training pipeline in which

the parameters of the BCI model are learn using user-

speciﬁc supervised machine learning models trained

on EEG recordings. The second module consists

on the online decoding system, which performs real-

time data streaming and inference using the model le-

arnt during the ofﬂine training process to decode the

player’s intention continuously.

2.1 Training Pipeline

The aim of the BCI Training pipeline is to learn the

parameters of a machine learning model for motor

imagery detection that will later be used for real-time

motor imagery decoding. These parameters are a set

of CSP ﬁlters (Ramoser et al., 2000) for feature ex-

traction and the parameters of a Logistic Regression

classiﬁer. The training pipeline is shown on Figure 2.

The system was trained on two healthy subjects who

had never participated in BCI studies or used any BCI

system before - one of them (LG) was a twenty one

year-old left-handed male and the other one (EG) was

a twenty three year-old right-handed male. Informed

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics

consent was obtained and the study was performed ac-

cording to institutional standards for the protection of

human subjects. In this section we describe the details

of the training pipeline.

Figure 2: BCI model training: Labeled EEG recordings of

motor imageries are fed to the training pipeline. The sig-

nals are band-pass ﬁltered between 7 and 30 Hz to capture

the mu and beta frequency bands associated to sensorimotor

rhythms. The band-pass ﬁltered data is used to ﬁt the Com-

mon Spatial Patterns, obtaining pairs of ﬁlters, such that one

maximizes the variance of one speciﬁc motor imagery and

minimizes the variance of another one, and the other ﬁlter

does the opposite. The logarithm of the variance of the spa-

tially ﬁltered sources are used as the features to ﬁt a Logistic

Regression classiﬁer. Variations of the model were perfor-

med using 2-4 motor imageries to classify.

Experimental Setup for Training Data Acquisition

The subject sits in a comfortable chair at approxima-

tely 70 centimeters away from a computer screen. The

EEG signal is recorded using a Brain Products’ Easy-

Cap with 32 Ag-AgCl active actiCAP electrodes loca-

ted according to the International 10-20 system. High

viscosity electrolyte gel for active electrodes is used

to reduce the impedance between the electrode and

the scalp under a tolerance level of 50 KOhm. 5 elec-

tromyogram (EMG) bipolar pairs of electrodes are lo-

cated on the subject’s right and left forearm (around

the extensor retinaculum area), on the right and left

feet (around the extensor retinaculum area) and on the

torso (around the rectus abdominis area). One pair of

bipolar electrooculogram (EOG) electrodes are loca-

ted around the vertical vEOG up and vEOG down lo-

cations. Figure 3 shows the experimental setup. The

BCI user watches game-play videos of the game and

executes motor thoughts when the avatar reaches dif-

ferent sections of the race track. The motor imageries

consist on imaginations of opening and closing the

right hand, opening and closing the left hand, ﬂexing

and relaxing both feet at the ankle (dorsiﬂexion) and

contracting and relaxing the abdominal muscles and

their associated game events are detailed on Figure 1.

This way we are able to gather and label motor ima-

gery data from a 32-electrode EEG cap. Note that for

this study, we used the ﬁrst version of the Cybathlon

Brain Runners videogame, which differs from the ﬁ-

nal one used in the Cybathlon competition. In order

to homogenize the number of pads of each class, the

ﬁve videos that maximized the homogeneity on the

number of pads of each type were selected, yielding

a total number of 90 trial events (22 green, 23 pur-

ple, 23 yellow and 22 grey). Each of these ﬁve videos

was repeated four times during the experiment, thus,

a total of 20 videos were played containing a total of

360 trial events (88 green, 92 purple, 92 yellow and

88 grey). Performing motor imageries is a tiring task,

so after each trial, the subject is allowed to rest and a

black screen appears informing him how many videos

he has already completed and asking him to press a

key to continue with the new video whenever he feels

prepared.

Figure 3: Experimental setup for training data acquisition

(the ﬁrst author shown using the systems in the photograph).

The subject sits on a comfortable chair with the 32 active

electrodes cap recording his EEG signal, 5 pairs of bipolar

EMG recording his muscle activity and one pair of bipolar

EOG electrodes recording his eye movements and blinks.

Training the BCI Model

After collecting and labelling training data for the

different motor imageries for a given user, we feed

it to the training pipeline. Our approach focuses on

spectral and spatial ﬁltering, using a band-pass ﬁlter

and learnt Common Spatial Patterns respectively

in event-locked time windows in order to capture

Event-Related-Desynchronisation (ERD) and Event-

Related-Synchronisation (ERS) features associated

to the motor imageries (Neuper et al., 2006). Once

the features of interest are extracted through spectral

and spatial ﬁltering, we use them to ﬁt a Logistic

regression classiﬁer to discriminate across the diffe-

rent motor imageries. For clarity reasons, this paper

ﬁrst describes the training platform to decode two

motor imageries and later on, an adaptation for for

multi-class classiﬁcation. The training process as

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race

shown on Figure 2, consists on band-pass ﬁltering,

selecting a time-window, ﬁtting the CSP ﬁlters, spa-

tially ﬁltering the signal using the learnt CSP ﬁlters,

extracting log-variance features and ﬁtting a classiﬁer.

Band-pass Filtering and Time-window Selection

The ﬁrst step on the training process is band-pass ﬁl-

tering the labeled EEG recordings. ERD and ERS

on the sensorimotor cortex are features observed on

EEG signals associated with motor imageries that are

prominent on the µ and β frequency bands. In order

to capture these events, the EEG recording is band-

pass ﬁltered using a Finite Impulse Response (FIR)

ﬁlter with cutoff frequencies between 7 and 30 Hz and

using a pass-band ripple of -20 dB and a stop-band

ripple of -40 dB.

ERD/ERS are shown to be active from 0.5

seconds after the stimulus onset and last between

2 and 3 seconds (Neuper et al., 2006). Thus, the

EEG chunk of interest for our motor imagery control

purpose is contained within this time-window. We

were also interested on building a BCI system that

reacts as fast as possible to the user’s intent, for

online decoding purposes. To achieve this, ideally,

we would use the shortest possible time-window

while keeping enough information on the EEG signal

to provide a reliable decoding. Considering this,

we selected a time window from 0.5 seconds after

the motor imagery onset until 2.5 seconds after the

motor imagery onset, obtaining multiple samples of

2-seconds long labeled time-series to train the system

capturing ERD/ERS events.

Learning Optimal CSP Filters

Common Spatial Patterns (Ramoser et al., 2000) are

pairs of ﬁlters that linearly transform the multichannel

time series of an EEG signal into lower-dimensional

time series in such a way that the ﬁrst ﬁlter maximizes

the voltage variance across time of signals of Class A,

while minimizing the variance of signals of Class B

and the second ﬁlter does the opposite, it minimizes

the variance of signals of Class A and maximizes the

variance of signals of class B. Class A and B are two

different motor imageries.

We show this in Figure 4, on a left-hand versus

right hand motor imagery trial. If we observe the

voltage distribution on a given trial on the C3 and C4

EEG channels, located on the left and right hemisp-

here of the motor cortex respectively (Figure 4A), we

can observe that the voltage distribution of both clas-

ses have an important amount of overlap and thus, the

variance of both clusters on each channel would not

be a good feature to separate the different-class trials

using a linear classiﬁer. After ﬁltering with a pair of

learnt spatial ﬁlters on Figure 4B, the signals obtai-

ned for the ﬁrst source S1 (after ﬁltering with the ﬁrst

CSP ﬁlter 1), are maximal in variance for the blue

cluster, corresponding to left hand movement and mi-

nimal for the red cluster, corresponding to right hand

movement. The opposite happens with the second

source S2 (after ﬁltering with the converse CSP ﬁl-

ter 2). In this case the red cluster, corresponding to

right hand movement event has maximal variance and

the blue cluster, corresponding to left hand movement

have minimal variance. The signals are now much

more linearly separable by extracting variance-based

features.

Figure 4: A) The distribution of a raw EEG signal of a right

hand trial (red) and left hand trial (blue) before the Com-

mon Spatial Filters are applied. B) Same trials after spatial

ﬁltering using a single pair of optimized CSP ﬁlters.

The supervised Machine Learning algorithm to learn

the CSP ﬁlters are is described as the following gene-

ralized eigenvalue problem, using the left versus right

hand motor imageries for illustration purposes:

We compute the average covariance matrices for each

class over all trials of our band-pass ﬁltered 2-second

long time-series training dataset, obtaining Σ

and

, for left and right hand classes respectively, and

ﬁnd the simultaneous diagonalizer W of Σ

and Σ

W = D

(1)

Such that the diagonal matrices D

and D

, fol-

low:

+ D

= I

(2)

This formulates a generalized eigenvalue problem

of the form:

W = λΣ

(3)

This is satisﬁed for W consisting of the generalized

eigenvectors w

( j=1,, C) and λ being: λ = λ

/λ

and λ

(c ∈ {L,R}), the diagonal elements of D

and

, respectively. It is important to note that λ

≥ 0,

is the variance of class c in the CSP source j, and that

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics

+ λ

= 1. Thus, a value λ

close to one, means

that the spatial ﬁlter w

yields a high variance in the

left hand class and a low variance on the right hand

class and the opposite if it is close to zero. This makes

both classes easily discriminative by computing the

variance across time on the CSP ﬁltered time series.

Once we obtain the eigenvectors matrix, we take

the n ﬁrst and the n last columns of W (n = 3 in our

case), as they are maximally informative to extract va-

riance features. The pair of ﬁlters composed by the

ﬁrst and last columns are the ones that provide max-

imal discriminative information between the variance

of the two classes, but the rest n − 1 pairs provide ro-

bustness for classiﬁcation.

Since the Common Spatial Patterns are vectors of

dimension equal to the number of EEG channels, they

can be visualized using bipolar plots, which provide

useful information regarding the EEG channels that

are enhanced or suppressed to provide maximal

discrimination between classes and gives us a good

intuition about which areas of the brain the CSP

ﬁlter pairs are trying to extract information from.

An example of the bipolar plot of a pair of common

spatial ﬁlters optimized for subject EG2608’s right

hand versus left hand data is shown on Figure 5.

CSP Pattern 1 shows that the ﬁlter is amplifying the

depolarization associated with the activation of the

left hemisphere on the motor cortex, which is known

to be related to right hand movements. CSP Pattern 2,

in contrast, ampliﬁes the depolarization of a similar

area of the motor cortex but in the right hemisphere,

associated to left hand movements. The visualization

of these ﬁlters, thus, provide a good interpretation

on whether the CSP ﬁlters are capturing ERD/ERS

features or other irrelevant EEG patterns, such as

artifacts, in which case we would observe a strong

weighting on the frontal electrodes. The visualization

of spatial ﬁlters also provide with an estimation on

which parts of the brain are being activated associated

to different motor imageries, which can vary across

different subjects and can be taken into account to

decide which motor imageries a speciﬁc subject

should use to control the BCI video game.

Fitting a Logistic Regression Classiﬁer

After ﬁtting the CSP ﬁlters, we extract variance featu-

res on the spatially ﬁltered signal that are informative

of the underlying motor imagery.

First, we ﬁlter all time series of each class with

the spatial ﬁlters. An EEG time-series is X ∈ R

t×c

being t, the number of time steps contained on the 2

seconds epoch, (t = 1000 in our case, as signals are

sampled at 500 Hz) and c the number of channels of

Figure 5: Bipolar plot of the ﬁrst pair of Common Spatial

Patterns from subject EG2608 optimized for the classiﬁca-

tion between right hand (CSP Pattern 1) and left hand (CSP

pattern 2). CSP Pattern 1 shows that the ﬁlter is strongly

weighting the left hemisphere of the motor cortex, associa-

ted to right hand movements. The second ﬁlter of the pair,

CSP Pattern 2, in contrast, strongly weights the the right

hemisphere of the motor cortex, associated to left hand mo-

vements. The colour map represents the weights assigned

to each of the channels by the spatial ﬁlters, dark blue being

the highest and red being the lowest. Bipolar plots crea-

ted using BCILAB (Kothe and Makeig, 2013) using a cubic

interpolation of W

−1

the EEG signal (31, in our case). Our learnt CSP ﬁl-

ters are, W ∈ R

t×2n

, being n the number of pairs of

ﬁlters being used (columns of the eigenvector matrix),

which is n = 3 in our case. Thus, each ﬁltered time-

series S ∈ R

t×2n

becomes:

S = XW

(4)

We then extract features from the spatially ﬁltered sig-

nals, which are discriminative with respect to their va-

riance. Thus, we build a feature space based on the

logarithm of the variance of S across time. The loga-

rithm compresses the range and reduces the impact of

outliers.

The features f ∈ R

1×2n

, are thus:

f = log(var(S))

(5)

We ﬁnally used the extracted features to train the L2-

regularized Logistic Regression classiﬁer using BCI-

LAB (Kothe and Makeig, 2013), which is built upon

the LIBLINEAR (Fan et al., 2008) and CVX (Grant

and Boyd, 2008) packages.

In order to evaluate the classiﬁcation perfor-

mance, we ran a 5-fold cross-validation on the

training dataset. Importantly, to avoid correlation

between nearby trials, we performed a block-wise

cross validation leaving 5 trials out between the

training and test blocks. Doing so attempts to remove

correlated trials between training and test sets on

each cross-validation split to avoid unrealistic ofﬂine

decoding accuracy results.

Multi-class Motor Imagery Classiﬁcation

CSP generates optimized spatial ﬁlters to discriminate

between two motor imageries by linearly transfor-

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race

ming the EEG signal to maximize or minimize the va-

riance of one of the pair of motor imageries. Despite

its binary nature, it can be extended to work with more

than two classes by computing the ﬁlters for all pos-

sible pairs of classes and ﬁtting classiﬁers for all pos-

sible pairs. We then compute the normalized probabi-

lity of each class across all binary Logistic Regression

classiﬁcation tasks and select the class with maximum

probability.

As an alternative, we also developed a two-step hier-

archical binary classiﬁer. For the ﬁrst step, we ﬁt a

model to classify the motor imagery associated to the

no-command signal against all the rest of motor ima-

geries. The second step is the standard classiﬁcation

problem as described before for the remaining motor

imageries.

The decoding then starts on the ﬁrst step, using a

Logistic Regression classiﬁer. If the rest of the classes

wins against the no-command class, then the system

enter the second decoding step to classify which is

the most likely motor imagery out of the rest of the

classes using a Logistic Regression classiﬁer.

2.2 Real-time BCI Module

Once the parameters of the BCI model have been

estimated running the training pipeline, they are used

to decode the mental state of the BCI user in real

time and send commands to control the video game

accordingly. Figure 6 summarizes the transition

from the real-time acquisition of EEG data to the

translation of this information into a videogame

command. In this section we describe the details of

real-time BCI module.

From an EEG Signal to a Video Game Command

In this section, we describe the speciﬁc case for two

classes and a Logistic Regression classiﬁer is explai-

ned, although variations using more than two classes

were also implemented.

We used a modiﬁed version of the RDA Server for

the BrainVision Recorder, streaming the data directly

from the PyCorder recording software into MATLAB.

A 10 seconds ring buffer was implemented and upda-

ted every 50 milliseconds.

The 10 seconds-long RDA server buffer is ﬁlte-

red using a Finite Impulse Response (FIR) ﬁlter with

cutoff frequencies between 7 and 30 Hz and using a

pass-band ripple of -20 dB and a stop-band ripple of

-40 dB. Once the buffer is band-pass ﬁltered, just the

last two seconds are used for decoding, matching the

time window length used to train the BCI model. The

reason for using a 10 seconds buffer for ﬁltering is

Figure 6: Real-Time Decoding BCI module. A chunk of

EEG data is streamed into a buffer of 10 seconds length

every 50 milliseconds. This buffer is accessed by the deco-

ding thread, which ﬁrst applies a band-pass ﬁlter between

7 and 30 Hz. Then, it takes the most recent 2 seconds of

the buffer and applies the CSP ﬁlters computed on the trai-

ning pipeline to it, obtaining a set of spatially ﬁltered sig-

nals. The feature vector is obtained by computing the log-

variance of the signals across time. The log-variance feature

vector is classiﬁed using the Logistic Regression classiﬁer.

The BCI system ﬁnally sends the corresponding command

to the video game system via UDP Protocol.

that the band-pass ﬁlter transients might distort the

ﬁrst time steps of the signal, but the last two seconds,

which are the part we are interested on, are free of

distortions.

The 31-channel, 2 seconds-long EEG signal is

spatially ﬁltered with the 3 pairs of CSP ﬁlters produ-

cing a time series of 2 seconds and 6 source channels.

The logarithm of the variance of each of these time

series is computed to produce a feature vector, which

is ﬁnally classiﬁed using the logistic regression clas-

siﬁer:

y =

1 + e

−(xw+b)

(6)

Being x ∈ R

1×2n

the log-variance feature, n the num-

ber of pairs of ﬁlters (3 in our case) and w and b the

parameters of the Logistic Regression Classiﬁer. y is

a probabilistic output informative about the probabi-

lity of the input signal belonging to class 1 (e.g Right

Hand motor imagery) and the converse 1 − y is the

probability of the input signal to belonging to class 2

(e.g. Left Hand motor imagery).

Finally, we map the classiﬁcation output to a

video game command. The video game has three

commands that execute different actions. In order to

send these commands to the video game, a speciﬁc

value has to be sent via UDP protocol to the speciﬁc

IP address of the game. In order to achieve a good

performance on the Cybathlon video game, it is im-

portant not to decode the wrong command, especially

during the grey pads within which no command must

be sent for a long period of time. For this reason,

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics

we adopted a conservative strategy to avoid false

positives and we only sent a command when a motor

imagery is decoded with a probability of 0.8 or higher

and a no-command otherwise.

Feedback to Improve Control of the System

The Real-Time BCI system provides feedback to the

user by displaying bars, representing the probability

of each of the motor imageries. This visual feedback

is very important in order for the user to be able to

adapt to the system. As in any natural learning task,

after some practice using the system, the user will be

able to adapt his brain activity to improve his skills

controlling the system.

A Brain-Computer interface is a system in which

both the computer and user’s brain activity must adapt

to each other to achieve an optimal performance. The

computer learns from the user through supervised ma-

chine learning and the user learns how to best control

the system through continuous training.

Once the users feel comfortable controlling the

BCI, they start using the system to control the Cy-

bathlon video game.

3 RESULTS

In the ﬁrst part of his section, we introduce our re-

sults evaluating our decoding accuracy ofﬂine, using

cross-validation on the acquired training data. In the

second part, we characterise the online decoding sy-

stem. Finally, we evaluate the real time decoding per-

formance.

3.1 Ofﬂine Classiﬁcation

We performed the described 5-fold block-wise cross-

validation to evaluate the training pipeline perfor-

mance. In this section we show the accuracy results

for all possible combinations of two and three motor

imagery decoding and the results for four motor ima-

gery decoding on two different subjects. Accuracy is

deﬁned as:

Accuracy(%) = (Correct/Total) × 100

(7)

For two motor-imagery classiﬁcation, the results on

all combinations of two pairs of motor imageries

using CSP with a Logistic Regression classiﬁer are

reported on this section. The bars represent the accu-

racy on the classiﬁcation of pairs of classes. The clas-

ses are labeled as S for thinking about contracting ab-

dominal muscles (stomach), RH for thinking about

opening and closing the right hand, LH for thinking

about opening and closing the left hand and F for thin-

king about expanding and contracting the feet. The

red line represents the chance level and the pink area

represents a 95% conﬁdence interval for the chance

level (M

uller-Putz et al., 2008). The blue bars repre-

sent the standard deviation for the classiﬁcation accu-

racy.

Results for subjects LG2807 and EG2608 are

shown on Figure 7. In the case of LG2807, our sy-

stem is able to differentiate between right hand and

feet with an accuracy of up to an 84.44%, however

it is unable to differentiate between some other motor

imageries, such as the case of stomach versus feet, for

which the classiﬁcation accuracy is within the chance

level range. For subject EG2608, our system is able

to decode three of the pairs of motor imageries with

up to an 81.67% accuracy for right hand versus left

hand classiﬁcation. However, it is unable to differen-

tiate between other pairs, such as the case of stomach

versus feet.

Similarly to the two-motor imagery case, the clas-

siﬁcation results using CSP with a Logistic Regres-

sion on all combinations of three motor imageries are

shown on Figure 8 and the classiﬁcation results for

the four motor imageries is shown in Figure 9. The

way in which the data is represented is equivalent to

the two motor imageries case.

The commands that subject LG2807 can best con-

trol for practical BCI purposes are feet versus right

hand, with an 84.44% accuracy for two classes, right

hand versus left hand versus feet with a 62.4% accu-

racy and for the four classes the decoding accuracy is

48.61%.

In the case of subject EG2608, the commands that can

be best controlled are right hand versus left hand with

an 81.67% accuracy for two classes, right hand versus

left hand versus feet with a 56.65% accuracy for three

classes and the decoding accuracy for the four classes

is 44.72%.

3.2 Real-time BCI Characterization

We describe the real-time BCI system characteristics

as:

Decoding delay: t

= 0.1214 ± 0.0031 seconds.

Decoding time-window length: t

= 2 seconds.

ERD/ERS delay: t

ERD/ERS

= 0.5 seconds.

The decoding delay refers to the minimum period of

time that it takes for our system to perform a full de-

coding task, this is, since the moment in which the

raw EEG data enters ours system until the moment

in which the decoded command is sent to the video

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race

Figure 7: Classiﬁcation accuracy between two motor ima-

geries for subjects LG2807 (top) and EG2608 (bottom)

using CSP with a Logistic Regression classiﬁer. Motor ima-

geries are labeled as: S = Stomach, RH = Right Hand, LH

= Left Hand, F = Feet. The red line represents chance label

and the pink area is a conﬁdence interval for chance level.

The blue bars represent the standard deviation for the clas-

siﬁcation accuracy. (These apply to all other ﬁgures).

Figure 8: Classiﬁcation accuracy between three motor ima-

geries for subject LG2807 (top) and EG2608 (bottom) using

CSP with a Logistic Regression classiﬁer. Using the same

legend described in Figure 7.

game system. This time is on average t

= 121 mil-

liseconds. The decoding response, however is limi-

ted by the length of the time window that we need to

use for the decoding. The EEG features that we are

using are relatively slow, at a maximum frequency of

Figure 9: Classiﬁcation accuracy between four motor ima-

geries for subject LG2807 (top) and EG2608 (bottom) using

the CSP with an Logistic Regression classiﬁer. Using the

same legend described in Figure 7.

30 Hz and even in the ideal case in which we would

be able to identify an ERD/ERS feature instantly, we

would still have a t

ERD/ERS

= 0.5s lag, which is an

intrinsic delay associated to these speciﬁc EEG featu-

res (Neuper et al., 2006). For training, we use a time

window between 0.5 seconds until 2.5 seconds after

the event onset, in order to capture ERD/ERS events.

For this reason, we need to keep a consistent in the

time-window length for real-time decoding, t

= 2s.

3.3 Real-time BCI Decoding Results

The real-time system was tested on one right-handed

subject, EG2608 using CSP with a Logistic Re-

gression classiﬁer using two and three commands

respectively. This was the second session for this sub-

ject. First, the subject completed the training phase

to optimise the CSP ﬁlters and Logistic Regression

classiﬁer for this second session. This time, however,

his ofﬂine performance was signiﬁcantly lower than

the during the ﬁrst session, obtaining a maximum

of 62.5% accuracy on two-class decoding for right

hand versus feet, a maximum of 52.29% accuracy for

three-class decoding for right hand versus left hand

versus feet. The subject then played the Cybathlon

game under three different conditions:

The ﬁrst condition consisted on two commands

decoding using a standard binary classiﬁcation.

Right hand motor imagery was assigned to control

the baseline command and stomach motor imagery

to control the accelerate command. The average

accuracy across 10 game rounds was: 68.62 %

The second condition consisted on three commands

using a three class classiﬁcation. Right hand motor

imagery was assigned to control the accelerate com-

mand, stomach motor imagery to control the jump

command and left hand motor imagery to control the

baseline command. The average accuracy across 3

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics

game rounds was: 47.66%

The third condition consisted on three commands

decoding using a two-step hierarchical classiﬁer.

Right hand motor imagery was assigned to control

the accelerate command and stomach motor imagery

was used to control the jump command. The ﬁrst

step of the hierarchical classiﬁer decides wether any

of the imageries are active and if the threshold is

greater than 0.8, it makes a prediction using the right

hand versus stomach classiﬁer. If any of the motor

imageries are decoded, its respective command is

sent, otherwise no command is sent. The average

online-decoding accuracy across 3 game rounds was:

60.03 %

The subject was not able to complete the online 4-

class decoding task as he struggled to achieve control

for 4 commands on the feedback phase. A snapshot of

the subject playing the Cybathlon video game using

motor thoughts is shown on Figure 10.

Figure 10: A subject plays the Cybathlon video game using

thoughts. The three bars at the right of the screen represent

the probability of a speciﬁc motor thought to be decoded. In

this snapshot the subject is successfully executing a jump.

4 DISCUSION

In this work we demonstrate how a data-efﬁcient

EEG-based online BCI system can achieve competi-

tive results on the Cybathlon BCI race task. The main

advantage of our system is that it requires a small

amount of training data, effectively translating in

short data acquisition sessions prior to live BCI usage

and potentially improving the BCI user experience.

Our system was able to achieve an average decoding

accuracy across subjects of 86%, 62% and 49%

for the ofﬂine classiﬁcation of two, three and four

motor imageries respectively, evaluated using a

correlation-corrected cross-validation approach. Our

system was also able to achieve an online decoding

accuracy of up to 60.03% controlling 3 commands,

while only requiring the acquisition of 360 motor

imagery trials.

Considerations Regarding CSP-based BCIs

Common Spatial Patterns have some drawbacks (To-

mioka et al., 2007). CSP were originally described as

a decomposition technique (Koles, 1991) rather than a

tool for classiﬁcation based on variance features. No-

netheless they have been extensively used for classiﬁ-

cation purposes in motor imagery based BCI applica-

tions with success.

One important problem, however, is that the si-

multaneous diagonalization of the covariance matri-

ces suffers greatly from the presence of even a small

amount of outlier trials on the EEG dataset, such as in

cases of artifact presence.

Moreover, the nature of the CSP ﬁlters is binary.

They attempt to maximize the difference in variance

between pairs of classes. They can be extended to

multi-class tasks as we described on the methods

section but they are not directly optimized for multi-

class classiﬁcation. Besides this workaround, other

alternatives to multi-class CSP have been attempted

in the past and might pose an appropriate choice for

the given problem (Grosse-Wentrup and Buss, 2008).

Previous Results on the Cybathlon BCI

Ortega et al. proposed an alternative decoding route

on their BCI for the Cybathlon using a Convolutional

Neural Network, consisting of a convolutional layer,

followed by a fully connected layer and a softmax

(Ortega et al., 2018a). They reported results for 4-

class decoding, using a 4-fold cross-validation, both

ofﬂine and online, obtaining up to 54.5% accuracy

ofﬂine, which shows that optimizing the data acqui-

sition approach to gather a larger dataset (9000 sam-

ples), allows deep neural networks to improve perfor-

mance compared to simpler linear classiﬁcations on

smaller datasets. Additionally, they reported 47% on-

line accuracy for 4-class decoding.

Schwartz et al. also showed relevant results on

their Cybathlon BCI system, using two CSP projecti-

ons, extracting 12 power-band features and ﬁtting a

shrinkage regularized Linear Discriminant Analysis

(Schwarz et al., 2016). They collected 1230 trials

and obtained a 66.1% median accuracy ofﬂine, on the

best combination of motor imageries using right hand

versus left hand versus feet versus rest, which inte-

restingly indicates a large improvement compared to

their results using four motor imageries. These results

demonstrate that including a rest state and gathering

more data can boost decoding performance. They also

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race

reported a drop in online performance for 4-class de-

coding, down to 51.2% accuracy across all user deci-

sions.

One of the reasons why our results may dif-

fer from the previous studies discussed above is the

choice of a different cross-validation method. In pre-

vious work, a standard block-wise cross-validation

was chosen. Instead, we attempted to propose a more

robust cross-validation technique, which corrects for

highly correlated contiguous trials between training

and test blocks by discarding the 5 trials in between

them for each split of our block-wise cross-validation.

Correlated trials between training and test sets might

be one of the factors involved in the ofﬂine to online

accuracy drop reported in previous work, but perhaps

not the only one. Note, that, even by using this ad-

vanced cross-validation technique to correct for cor-

related trials across training and test sets, we also

experienced an ofﬂine to online accuracy drop on 3-

classes decoding (although better results were obtai-

ned online than ofﬂine for the 2-class decoding case).

This suggests that there could be additional factors

involved in the ofﬂine to online accuracy drop. On

the whole our platform manages to offer correlation-

corrected cross-validation results for ofﬂine decoding,

which are at competitive levels of performance com-

pared to previous approaches, while signiﬁcantly re-

ducing the amount of training data needed to operate

the BCI (360 samples compared to thousands of sam-

ples employed in other studies). It thereby poses a so-

lid decoding system that is aimed at managing more

efﬁciently the trade-off between decoding accuracy

and the length of training sessions.

For online BCI decoding tasks other than the

Cybtahlon, previous studies have shown promising

results on time-locked tasks. Friedrich et al. achieved

accuracies for 4-class decoding ranging between 61%

and 72% on 14 subjects, where 8 of them achieved

performance over the chance level for all 4 classes,

including motor and non-motor imageries (Friedrich

et al., 2013). In the case of 3-class decoding, Milln

et al. showed accuracies ranging between 55% and

76% across 5 subjects, also including both motor and

non-motor imageries (Mill

an et al., 2004).

Comparing Ofﬂine and Online Results

Results on real time decoding accuracy for subject

EG2608 showed a classiﬁcation accuracy of 68.62%,

surpassing the ofﬂine cross-validation results for the

same subject and session (62.5%). After several ga-

mes, the subject was able to achieve good control

over the video game tracks where only two commands

were needed.

In the case of three commands, the online accu-

racy was 47.66%, showing a drop compared to the

52.29% accuracy obtained ofﬂine, evidencing that the

control of three commands is much more challenging.

We also demonstrated that 3-class accuracy can be

greatly improved by using a two-step classiﬁcation

scheme, which executes binary classiﬁcation on each

step, obtaining a 60.03% accuracy.

In the three commands case, the subject reported

that control was possible provided enough concentra-

tion but signiﬁcantly harder than the two commands

case. The subject also attempted 4-class decoding,

however, during the feedback phase, the subject was

unable to achieve meaningful control. It is worth noti-

cing that previous studies on BCIs applied to commu-

nication have shown that a minimum of 70% accuracy

is necessary for the system to be usable (K

ubler et al.,

2001). We consider improvements on the 4-class ga-

ming control as one of the primary outlook research

areas for the given decoding context and as a prioriti-

zed extension of the present version of our platform.

We also observed a high variability in perfor-

mance across the two subjects and across different

sessions, verifying that consistency and personali-

zation is an important challenge that needs to be

addressed for practical BCI adoption in clinical

settings.

Data Efﬁciency Assessment

Our approach builds upon previous work on ofﬂine

data-efﬁcient EEG BCI decoding (Ferrante et al.,

2015), where results for 2-class motor-imagery de-

coding using a CSP-based linear system showed an

average accuracy of 88% when trained on just a do-

zen of samples per-class and takes it one step further,

proving satisfying decoding performance for multi-

class motor imagery decoding, as well as suitability

for practical online BCI tasks.

In particular, our system was able to achieve com-

petitive results compared to previous Cybathlon sy-

stems using a number of training samples one order

of magnitude smaller. More precisely, we trained our

classiﬁer using 360 samples, achieving an ofﬂine 4-

class mean decoding accuracy across subjects of 49%

on our correlation-corrected cross-validation asses-

sment. When compared against the ofﬂine block-wise

cross-validation results from previous Cybathlon sy-

stems on the same task, one of them (Ortega et al.,

2018a) showed a 11% increase in accuracy at the ex-

pense of a 2400% increase in training data volume

and a second one (Schwarz et al., 2016) obtained a

35% accuracy increase at the expense of increasing

training data volume by 242%.

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics

The main characteristic making our system data-

efﬁcient is that it only needs to optimize a very small

set of model parameters, compared to other appro-

aches using much more complex non-linear models

(Ortega et al., 2018a; Yang et al., 2015; Tabar and

Halici, 2017; Lu et al., 2017). In fact, our classiﬁer

only needs to learn 6 weights and one bias term.

Overall, our system aims to ﬁnd an optimal trade-

off between accuracy and volume of training data nee-

ded prior to BCI live usage, so as to obtain satisfying

levels on the former, while reducing the latter dras-

tically, effectively shortening the duration of training

sessions. This could potentially lead to an improved

BCI user experience, enabling mainstream adoption

of such systems in the clinical setting.

5 CONCLUSION

In this work we demonstrate how a data-efﬁcient

EEG-based online BCI system can be deployed end-

to-end to compete in the Cybathlon BCI race. The

main advantage of our system compared to others

from previous literature is that it requires a volume

of training data one order of magnitude smaller while

retaining a competitive accuracy, essentially reducing

the amount of time needed for data acquisition prior

to live BCI usage. In addition, we present methods

to troubleshoot an EEG-based BCI system by exa-

mining CSP bipolar plots, pre-CSP and post-CSP ﬁl-

tered trials. We also demonstrate good practices to

avoid over-optimistic ofﬂine results by using a block-

wise cross-validation discarding a number of trials be-

tween the training and test blocks, effectively elimina-

ting correlated trials.

The system we implemented was based on tem-

poral ﬁltering using band-pass ﬁltering and spatial ﬁl-

tering using Common Spatial patterns to extract log-

variance features to ﬁt a linear classiﬁer (Logistic Re-

gression). We focused on the identiﬁcation of senso-

rimotor rhythms associated with motor imageries and

4 different motor imageries were tested (right hand

movement, left hand movement, both feet movement,

and abdominal contraction). Our results showed com-

parable results in ofﬂine and online accuracy for 2-

commands control, obtaining superior results on the

online mode than on the ofﬂine mode, but a drop

in performance for 3-class online decoding with re-

spect to ofﬂine decoding. We were able to improve

3-command online control by using a two-step binary

classiﬁcation strategy.

We also observed great variability in performance

across subjects and sessions, which suggests that mul-

tiple sessions would be ideally needed to identify

which motor imageries can be best controlled by each

subject, so that training can be focused on those spe-

ciﬁc ones. The data-efﬁcient nature of our approach

makes it well suited for fast identiﬁcation of optimal

motor-imageries per subject. Finally, the Cybathlon

testing task and performance conditions of our sy-

stem align it with contemporary assistive technolo-

gies’ standards and afford it the potential to become

a good benchmark and a promising tool for the ad-

vancement of non-invasive BCI research and applica-

tions.

REFERENCES

Dziemian, S., Abbott, W. W., and Faisal, A. A. (2016).

Gaze-based teleprosthetic enables intuitive continu-

ous control of complex robot arm use: Writing dra-

wing. In 2016 6th IEEE International Conference on

Biomedical Robotics and Biomechatronics (BioRob),

pages 1277–1282.

Fan, R., Chang, K., Hsieh, C., Wang, X., and Lin, C. (2008).

LIBLINEAR: A library for large linear classiﬁcation.

Journal of Machine Learning Research, 9:1871–1874.

Ferrante, A., Gavriel, C., and Faisal, A. A. (2015). Data-

efﬁcient hand motor imagery decoding in EEG-BCI

by using morlet wavelets and common spatial pattern

algorithms. In 2015 7th International IEEE/EMBS

Conference on Neural Engineering (NER), pages

948–951.

Friedrich, E. V., Scherer, R., and Neuper, C. (2013). Long-

term evaluation of a 4-class imagery-based braincom-

puter interface. Clinical Neurophysiology, 124(5):916

– 927.

Grant, M. C. and Boyd, S. P. (2008). Graph implemen-

tations for nonsmooth convex programs. In Recent

Advances in Learning and Control. Lecture Notes in

Control and Information Sciences, volume 371, pages

95–110, London. Springer.

Grosse-Wentrup, M. and Buss, M. (2008). Multiclass com-

mon spatial patterns and information theoretic feature

extraction. IEEE Transactions on Biomedical Engi-

neering, 55(8):1991–2000.

Holz, E. M., Botrel, L., Kaufmann, T., and K

ubler, A.

(2015). Long-term independent brain-computer inter-

face home use improves quality of life of a patient in

the locked-in state: A case study. Archives of Physical

Medicine and Rehabilitation, 96(3):S16–S26.

Koles, Z. (1991). The quantitative extraction and topo-

graphic mapping of the abnormal components in the

clinical EEG. Electroencephalography and Clinical

Neurophysiology, 79(6):440 – 447.

Kothe, C. A. and Makeig, S. (2013). BCILAB: a platform

for braincomputer interface development. Journal of

Neural Engineering, 10(5):056014.

ubler, A., Neumann, N., Kaiser, J., Kotchoubey, B., Hin-

terberger, T., and Birbaumer, N. P. (2001). Brain-

computer communication: Self-regulation of slow

Data-efﬁcient Motor Imagery Decoding in Real-time for the Cybathlon Brain-Computer Interface Race

cortical potentials for verbal communication. Ar-

chives of Physical Medicine and Rehabilitation,

82(11):1533–1539.

Lu, N., Li, T., Ren, X., and Miao, H. (2017). A deep le-

arning scheme for motor imagery classiﬁcation based

on restricted boltzmann machines. IEEE Transacti-

ons on Neural Systems and Rehabilitation Engineer-

ing, 25(6):566–576.

Maimon-Mor, R. O., Fernandez-Quesada, J., Zito, G. A.,

Konnaris, C., Dziemian, S., and Faisal, A. A. (2017).

Towards free 3D end-point control for robotic-assisted

human reaching using binocular eye tracking. In 2017

International Conference on Rehabilitation Robotics

(ICORR), pages 1049–1054.

Makin, T. R., de Vignemont, F., and Faisal, A. A. (2017).

Neurocognitive barriers to the embodiment of techno-

logy. Nature Biomedical Engineering, 1:0014 EP.

Milln, J. R., Renkens, F., Mourio, J., and Gerstner, W.

(2004). Brain-actuated interaction. Artiﬁcial Intelli-

gence, 159(1):241 – 259.

uller-Putz, G. R., Scherer, R., Brunner, C., Leeb, R., and

Pfurtscheller, G. (2008). Better than random: a closer

look on BCI results. International Journal of Bioelec-

tromagnetism, pages 52–55.

Neuper, C., Wrtz, M., and Pfurtscheller, G. (2006).

ERD/ERS patterns reﬂecting sensorimotor activation

and deactivation. In Neuper, C. and Klimesch, W.,

editors, Event-Related Dynamics of Brain Oscillati-

ons, volume 159 of Progress in Brain Research, pages

211 – 222. Elsevier.

Noronha, B., Dziemian, S., Zito, G. A., Konnaris, C., and

Faisal, A. A. (2017). Wink to grasp - comparing eye,

voice and EMG gesture control of grasp with soft-

robotic gloves. In 2017 International Conference on

Rehabilitation Robotics (ICORR), pages 1043–1048.

Obermaier, B., Neuper, C., Guger, C., and Pfurtschel-

ler, G. (2001). Information transfer rate in a ﬁve-

classes brain-computer interface. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

9(3):283–288.

Ortega, P., Colas, C., and Faisal, A. A. (2018a). Compact

convolutional neural networks for multi-class, perso-

nalised, closed-loop EEG-BCI. In 2018 IEEE Inter-

national Conference on Biomedical Robotics and Bi-

omechatronics (BioRob), 7th IEEE International Con-

ference (forthcoming).

Ortega, P., Colas, C., and Faisal, A. A. (2018b). Convoluti-

onal neural network, personalised, closed-loop brain-

computer interfaces for multi-way control mode swit-

ching in real-time. bioRxiv.

Ramoser, H., Muller-Gerking, J., and Pfurtscheller, G.

(2000). Optimal spatial ﬁltering of single trial EEG

during imagined hand movement. IEEE Transactions

on Rehabilitation Engineering, 8(4):441–446.

Riener, R. and Seward, L. J. (2014). Cybathlon 2016.

In 2014 IEEE International Conference on Systems,

Man, and Cybernetics (SMC), pages 2792–2794.

Schwarz, A., Steyrl, D., and Mller-Putz, G. R. (2016).

Brain-computer interface adaptation for an end user

to compete in the Cybathlon. In 2016 IEEE Interna-

tional Conference on Systems, Man, and Cybernetics

(SMC), pages 001803–001808.

Tabar, Y. R. and Halici, U. (2017). A novel deep learning

approach for classiﬁcation of EEG motor imagery sig-

nals. Journal of Neural Engineering, 14(1):016003.

Tomioka, R., Aihara, K., and M

uller, K. (2007). Logis-

tic regression for single trial EEG classiﬁcation. In

Sch

olkopf, B., Platt, J. C., and Hoffman, T., editors,

Advances in Neural Information Processing Systems

19, pages 1377–1384. MIT Press.

Tostado, P. M., Abbott, W. W., and Faisal, A. A. (2016).

3D gaze cursor: Continuous calibration and end-point

grasp control of robotic actuators. In 2016 IEEE In-

ternational Conference on Robotics and Automation

(ICRA), pages 3295–3300.

Yang, H., Sakhavi, S., Ang, K. K., and Guan, C. (2015).

On the use of convolutional neural networks and aug-

mented CSP features for multi-class motor imagery

of EEG signals classiﬁcation. In 2015 37th Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 2620–

2623.

Zhang L., He W., H. C. and Wang, P. (2010). Improving

mental task classiﬁcation by adding high frequency

band information. Journal of Medical Systems, 34:51

– 60.

NEUROTECHNIX 2018 - 6th International Congress on Neurotechnology, Electronics and Informatics