Finding the Optimal Time Window for Increased Classiﬁcation Accuracy

during Motor Imagery

D. A. Blanco-Mora

1 a

, A. Aldridge

1 b

, C. Jorge

1 c

, A. Vourvopoulos

3 d

, P. Figueiredo

3 e

and

S. Berm

udez i Badia

1,2 f

Madeira Interactive Techonologies Institute, Universidade da Madeira, Funchal, Portugal

Faculdade de Ci

encias Exatas e da Engenharia, Universidade da Madeira, Funchal, Portugal

Institute for Systems and Robotics - Lisboa, Instituto Superior T

ecnico, Universidade de Lisboa, Lisbon, Portugal

Keywords:

Brain-computer Interface, BCI, Motor Imagery, MI, Classiﬁcation Accuracy, Common Spatial Pattern, CSP,

Electroencephalography, EEG, Neurorehabilitation, Stroke.

Abstract:

Motor imagery classiﬁcation using electroencephalography is based on feature extraction over a length of

time, and different conﬁgurations of settings can alter the performance of a classiﬁer. Nevertheless, there

is a lack of standardized settings for motor imagery classiﬁcation. This work analyzes the effect of age on

motor imagery training performance for two common spatial pattern-based classiﬁer pipelines and various

conﬁgurations of timing parameters, such as epochs, windows, and offsets. Results showed signiﬁcant (p

≤ 0.01) inverse correlations between performance and feature quantity, as well as between performance and

epoch/window ratio.

1 INTRODUCTION

In recent decades, Brain-Computer Interfaces (BCIs)

have been used in novel neurorehabilitative tech-

niques, yielding promising results in terms of motor

recovery (Cervera et al., 2018). The use of BCIs in

neurorehabilitation allows the recruitment and activa-

tion of motor regions through Motor Imagery (MI),

without the need of active movement. This could

potentially result in neuro-plasticity changes in ar-

eas considered to be damaged from a stroke (Bai

et al., 2020). When combined with serious games

and Virtual Reality (VR), BCIs allow for a more in-

tensive neurorehabilitation (Putze, 2019), encourag-

ing patients via immediate feedback (Mubin et al.,

2020) and immersing them in engaging, virtual en-

vironments (Khan et al., 2020). One important chal-

lenge with neurorehabilitation is the timing and efﬁ-

cacy of feedback delivery relating to MI. Feedback

delivery plays an important role in BCIs, and efﬁ-

https://orcid.org/0000-0003-2232-0999

https://orcid.org/0000-0003-3733-4736

https://orcid.org/0000-0002-7693-7292

https://orcid.org/0000-0001-9676-8599

https://orcid.org/0000-0002-0743-0869

https://orcid.org/0000-0003-4452-0414

ciently using proprioceptive feedback can improve

BCI performance signiﬁcantly (Ramos-Murguialday

et al., 2019). During learning, feedback provided to

close the sensorimotor loop should be associated with

a MI event to facilitate the recreation of a more re-

alistic and authentic sensation. Similarly, feedback

provided during online sessions should be delivered

as soon as possible in response to a MI event.

The practice of detecting and classifying MI can

be challenging due to the variability and uniqueness

of each EEG signal. According to (Ortner et al.,

2015), 35% of participants obtained a classiﬁcation

accuracy lower than 70% with only 65% of people

controlling MI-based BCI adequately (≥ 70%). The

inability to replicate a standard pattern or brain net-

work topology, known as illiteracy, has a negative im-

pact on MI performance classiﬁcation (Ahn and Jun,

2015). To face this challenge, researchers use strate-

gies for the recruitment of participants to help bypass

BCI illiteracy (Ahn et al., 2013), but this is not an op-

tion when using BCI as a neurorehabilitative tool for

stroke patients, because the impaired neural pathways

prevent the use of screening strategies based on BCI

illiteracy, and the use of screening strategies will re-

strict the usability impact to a particular group with a

certain type of lesion.

144

Blanco-Mora, D., Aldridge, A., Jorge, C., Vourvopoulos, A., Figueiredo, P. and Bermúdez i Badia, S.

Finding the Optimal Time Window for Increased Classiﬁcation Accuracy during Motor Imagery.

DOI: 10.5220/0010316101440151

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 1: BIODEVICES, pages 144-151

ISBN: 978-989-758-490-9

Another way researchers are tackling the chal-

lenges of MI detection and classiﬁcation is by us-

ing machine learning tools to create methodologies

or pipelines that can be generalized across all users.

The most common methodologies use frequency

features or sensory motor rhythms (SMR), includ-

ing event-related synchronization/desynchronization

(ERS/ERD) (Padﬁeld et al., 2019). Common Spa-

tial Patterns (CSPs) algorithm is a feature extrac-

tion method that can learn spatial ﬁlters maximizing

the discriminability of two classes. CSP ﬁlters are

widely used to reduce data dimensionality for lighter

data processing and maximize the discriminability of

classes, for instance, right versus left arm movement

or arm versus leg movement. They also express data

as a combination of a reduced number of components

(Ai et al., 2018). Combined with linear discrimi-

nant analysis (LDA), CSP and LDA, are the preferred

pipeline and especially useful during online rehabili-

tation sessions (Lotte et al., 2018).

Even with improved techniques and methodolo-

gies, working with MI is challenging because MI clas-

siﬁcation is difﬁcult to generalize and to improve. For

instance, studies have performed different time set-

tings for signal analyses (Chen et al., 2020; Wang

et al., 2020), and it is known that ERS/ERD could last

for several seconds following the stimulus presenta-

tion (Pfurtscheller and Lopes, 1999). Hence, there is

a lack of standardized settings for MI due to the vari-

ability seen across subjects (Wang et al., 2020). Most

BCI research is done with healthy, young participants

whereas the stroke population tends to be elderly or

aged. Therefore, we designed a study to investigate if

our ﬁndings generalize to an older population.

In the context of a serious game developed for mo-

tor rehabilitation based on MI, we studied the effect of

age on MI signals and identiﬁed the optimal pipeline

and time parameter settings in comparison to a control

group of healthy individuals.

2 METHODS

To ﬁnd the optimal conﬁguration and time parameter

settings, we have designed a set of different conﬁg-

urations and tested them with NeuRow, a BCI train-

ing paradigm in VR, designed for upper-limb motor

rehabilitation (Vourvopoulos et al., 2016). The fol-

lowing subsections describe the software, equipment,

data acquisition, and processing pipelines we used for

this study.

2.1 NeuRow

NeuRow uses MI to control avatar movement and

haptic feedback in order to increase the sense of em-

bodiment of the user in a closed visual and senso-

rimotor loop (Vourvopoulos et al., 2019). With the

use of a head-mounted or a desk display, the game

shows the player in a ﬁrst-person perspective as an

avatar in a kayak ﬂoating on a body of water. Neu-

Row consists of two stages: training and real-time

control. The training stage uses signal processing

and signal features to train a system to separate EEG

signals according to desired events, speciﬁcally right

and left hand MI. The MI training stage is an adap-

tation of the Graz paradigm, using directional arrows

(Pfurtscheller et al., 2003). Participants perform MI

while observing the NeuRow avatar row to the left

or right for a total of 20 trials per side, following the

training block paradigm shown in Fig. 1.

Figure 1: Training Block paradigm.

2.2 Setup

The full setup includes a desktop computer, the Neu-

Row game, an Oculus Rift VR system, the EEG sys-

tem, and the custom-made haptic feedback system de-

scribed in (Vourvopoulos et al., 2016), as is shown in

Fig. 2. OpenVibe (Renard et al., 2010) was used to

translate MI events into game commands for control-

ling the NeuRow avatar via a Virtual-Reality Periph-

eral Network. Four stages constitute MI translation:

data acquisition, spatial ﬁlter training, classiﬁer train-

ing, and online game-play.

2.3 Participants

All participants are healthy with no known neurolog-

ical clinical history. The participants have been re-

cruited based on their motivation to participate in the

study and divided in two groups: the control group

and the age-matched group. To study the applicability

Finding the Optimal Time Window for Increased Classiﬁcation Accuracy during Motor Imagery

145

Figure 2: NeuRow setup: A) EEG system, B) The Oculus

Rift DK1 Head-Mounted Display, C) Haptic feedback sys-

tem, D) NeuRow BCI-VR task (Vourvopoulos, 2018).

of our ﬁndings for stroke participants and under the

approval of the ethical committee at the Hospital of

Madeira, N

elio Mendonc¸a (SESARAM), the spouses

of stroke patients were recruited as age-matched par-

ticipants to match the typical age of stroke survivors.

The control group consists of six young, healthy par-

ticipants with an average age of 25 ±5.33 years, 5

males and 1 female. The age-matched group consists

of 5 females and 1 male with an average age of 51.33

±4.97 years. All but one of the participants (one from

the control group) are right-handed according to the

Edinburgh inventory (Oldﬁeld, 1971).

2.4 Data Acquisition

EEG data was acquired using OpenVibe’s acquisi-

tion server and 32 EEG channels with a sampling

frequency of 500Hz and are downsampled to 250Hz.

The control participants’ datasets are acquired with

Liveamp 32 EEG ampliﬁer (Brain Products GmbH,

Munich, Germany) at IST in Lisbon, Portugal. The

age-matched participants’ datasets was acquired with

a wireless g.Nautilus from g.tec at the Madeira In-

teractive Technologies Institute in Funchal, Portugal.

Participants went through 20 trials per side doing MI

while watching the avatar. Common electrodes for

both groups: FC5, F3, Fz, F4, FC6, C3, Cz, C4,

CP5, CP1, CP2, CP6, P3, Pz, and P4. Age-matched

group included PO3, PO4, C1, and C2. Peripheral

electrodes (Fp1, Fp2, O1, O2, among others) are ex-

cluded from data processing to reduce the presence of

artifacts. The electrode impedances are checked using

the g.needAccess software from g.tec (g.tec medical

engineering GmbH, Austria).

2.5 Processing Pipeline

In the following, the term pipeline, as outlined in this

paper, refers to the processing structure used in the

classiﬁer training scenario. Two different processing

pipelines, both with the same LDA classiﬁer type, are

used for comparing classiﬁer performance in Open-

Vibe. These are referred to as 2-CSP-Boxcar and 4-

CSP-Hamming:

• The 2-CSP-Boxcar pipeline, applied in (Sury-

otrisongko and Samopa, 2015), has a classiﬁer

that is based on the power signal of a two-

dimensional CSP ﬁlter. It extracts power from

both output signals of the CSP ﬁlter during the

sliding window period, according to Equation (1)

in which x represents an output signal from the

CSP ﬁlter.

2 −CSP −Boxcar

f eature

= log(1 + x

) (1)

• The 4-CSP-Hamming pipeline, explained in

uller-Gerking et al., 1999) and applied in (Ir-

imia et al., 2018), uses the variance contribution

of each signal from a four-dimensional CSP ﬁl-

ter. It extracts the contribution of each of the four

CSP output signals, as shown in Equation (2) (Ir-

imia et al., 2018), with sub-index p representing

each output signal.

Hamming

f eature

= log

VAR

∑

p=1

VAR

(2)

EEG data processing for the two pipelines is exe-

cuted using CSP ﬁlter training and classiﬁer training

scenarios in OpenVibe. Raw EEG datasets are ﬁltered

from 8 to 30 Hz (including the Alpha and Beta bands)

in the respective CSP ﬁlter training scenarios and then

spatially ﬁltered using the previously trained CSP ﬁl-

ter for both pipelines. The CSP ﬁlter reduces the sig-

nal quantity obtained from the number of the EEG

channels to a ﬁxed number of representative outputs,

and it acquires weights for the EEG channels’ signals

according to their contribution to each event. Each

EEG channel’s signal is expressed as a linear combi-

nation of the representative outputs. The use of fre-

quency ﬁlters limits the signal content to frequencies

of interest, helping to denoise the desired signal and

eliminate potential constant offset, linear trending, or

noise that is caused by the power supply (50/60 Hz)

present in the signal.

The 2-CSP-Boxcar pipeline uses the ﬁltered data

to train the CSP ﬁlter to yield two output signals. The

4-CSP-Hamming pipeline, however, yields 4 output

signals. After the CSP ﬁltering, left and right MI tri-

als are windowed either with a Boxcar window or a

Hamming window, respectively. Next, EEG feature

extraction is applied to obtain the power and variance

characteristics to represent the signal. Finally, the ex-

tracted features from both pipelines are used to train

a LDA classiﬁer to identify right and left MI trials.

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

146

2.6 Pipeline Settings

To identify the optimal conﬁguration of the pipelines

described above, we study a number of settings. As

shown in Fig. 3, there are four main settings to con-

sider for the proposed pipelines: epoch length, epoch

offset time, sliding window length, and sliding win-

dow interval. Here, epoch length is the length of

the signal used for feature extraction; epoch offset

time determines the start of an epoch; sliding window

length is the length of the window that slides across

the selected epoch; and sliding window interval (mov-

ing steps) is the time between the start of each sliding

window. Sliding window displacement does not exist

when the sliding window has the same length as the

epoch. Relations between the sliding window length

and epoch length determine the quantity of data points

of the calculated or extracted EEG features as power

and variance contribution.

Figure 3: Parameters of a sliding window.

2.7 Data Analysis

The acquired EEG datasets are used to train the 2-

CSP-boxcar and 4-CSP-Hamming pipelines for vari-

ous combinations of time parameters. The classiﬁer’s

accuracy performance for each conﬁguration is com-

puted using 5-fold cross-validation, as the average of

the positive predictive value and the negative predic-

tive value from the confusion matrix’s classiﬁer per-

formance results. The association between the num-

ber of features used and the resulting accuracy per-

formance is tested using Pearson’s correlation. The

same is done for the relation between accuracy per-

formance and epoch/window ratios. Double-tailed t-

tests are used to test for signiﬁcant results.

The number of features’ data points (from now on

is shortened as only features) that feed the classiﬁer

depends on the epoch lengths, the window lengths,

the number of trials, and the number output channels

from the CSP ﬁlter, as determined by Equation (3).

f eatures





+ 1



+ 1



∗trials ∗CSP

out puts

(3)

In Equation (3), l is epoch length, w stands for

sliding window length, p is sliding window interval

or moving steps, k k denotes ﬂoor function, * means

product, trials represents the number of executed tri-

als during MI, and CSP

out puts

represents the number

of CSP signals.

3 RESULTS

To study the effects of age on MI and how pipeline

settings affect classiﬁer performance, the results sec-

tion is broken down into: an analysis of features

and their corresponding classiﬁcation performance,

pipeline performance analysis.

3.1 Features and Classiﬁer Performance

For the analysis of this section, we consider the

following variables: pipeline (2-CSP-Boxcar and 4-

CSP-Hamming), epoch length (1, 2, 4 s), sliding win-

dow length (0.5, 1, 2, 4 s), offset (0.1, 0.5 s), and

number of features (Equation (3)), keeping the sliding

window time interval as 50 ms. The combination of

these parameters totals 36 different conﬁgurations to

study. Fig. 4 shows the relationship between number

of features and the epoch and sliding window lengths.

The number of features is constant where the epoch

and sliding window lengths are equal. Hereafter,

these conﬁgurations are mentioned as epoch-lagged

because they add a time lag. The number of features

increases for longer epoch lengths and shorter sliding

window lengths. Conﬁgurations that have longer slid-

ing window lengths than epoch lengths are indicated

as NaN as it is not possible to have a negative quantity

of features.

Figure 4: Number of features per CSP output signal

(CSP

out puts

= 1) for varying epoch and sliding window

lengths (diagonal lines represent epoch-lagged conﬁgura-

tions).

The relationship between number of features and

classiﬁer performance was tested for correlation, ﬁnd-

Finding the Optimal Time Window for Increased Classiﬁcation Accuracy during Motor Imagery

147

ing Inverse relations between number of features and

classiﬁer performance for both pipelines. The anal-

ysis revealed correlations for all conﬁgurations to be

between -0.73 and -0.95 (Table 1). Interestingly, we

found a higher correlation between the number of fea-

tures for the age-matched group than for the con-

trol one of about 8%. When comparing the contri-

butions of different pipeline settings (epoch length

and sliding window length) to the classiﬁcation per-

formance, the correlation analysis revealed that the

epoch length/sliding window length ratio highly cor-

relates with the quantity of used features (r=0.8712,

p=0.002) and performance (-0.77 to -0.98) (Table 2).

Consistent with the previous data, correlations are

generally stronger for age-matched controls.

Table 1: Correlation values between classiﬁer performance

and quantity of features.

Offset = 0.1 s Offset = 0.5 s

r-value Boxcar Hamming Boxcar Hamming

Age-matched −0.94

∗∗

−0.93

∗∗

−0.95

∗∗

−0.94

∗∗

Controls −0.86

∗∗

−0.85

∗∗

−0.73

∗

−0.86

∗∗

∗p<0.05; ∗ ∗ p<0.01

Table 2: Correlation values between classiﬁer performance

and epoch/window length ratio.

Offset = 0.1 s Offset = 0.5 s

r-value Boxcar Hamming Boxcar Hamming

Age-matched −0.98

∗∗

−0.93

∗∗

−0.91

∗∗

−0.88

∗∗

Controls −0.92

∗∗

−0.77

∗

−0.82

∗∗

−0.92

∗∗

∗p<0.05; ∗ ∗ p<0.01

3.2 Pipeline Performance Analysis

3.2.1 Age-matched Participants

The age-matched group’s performance results show

the highest accuracy percentage of 85% for the 4-

CSP-Hamming pipeline (offset of 0.5 s, epoch and

window lengths of 1 s) (Fig. 5). The best results for all

conﬁgurations were achieved with equivalent epoch

and sliding window lengths, which as previously ex-

plained, are the conﬁgurations that add a signiﬁcant

time lag (1 to 4 s), making them unsuitable for appli-

cations that rely on real-time feedback. When we con-

sider only the non-epoch-lagged conﬁgurations that

are suitable for real-time feedback, the best perfor-

mance is seen for the 4-CSP-Hamming pipeline at

79.09% with an offset of 0.1 seconds, an epoch length

of 2 seconds, and a window length of 1 second.

Performances for the 2-CSP-Boxcar pipeline were

always lower than for the 4-CSP-Hamming pipeline,

both for epoch-lagged conﬁgurations (76.67% with an

offset of 0.5 s and epoch and window lengths of 4 s)

Figure 5: Classiﬁer performance for age-matched group (di-

agonal lines represent epoch-lagged conﬁgurations).

and for non-epoch-lagged ones (71.25% with an off-

set of 0.5 seconds, epoch length of 2 seconds, and

window length of 1 second).

3.2.2 Control Participants

Similar to the results of the age-matched group, the

control group’s best results were for the epoch-lagged

conﬁgurations. The highest performance was 90% for

the 4-CSP-Hamming pipeline when offset equaled 0.1

seconds. The best result for the non-lagged conﬁgu-

rations was also obtained with the 4-CSP-Hamming

pipeline at 79.68% for an offset of 0.1, an epoch

length of 1 second, and a window length of 0.5 sec-

onds (Fig. 6). For the 2-CSP-Boxcar pipeline, the best

lagged conﬁguration was recorded at 74.25% when

offset equaled 0.5 seconds, and the best non-lagged at

72.41% with 0.1 seconds of offset. Interestingly, there

seems to be a tendency for higher accuracies with

smaller epoch/sliding window ratios and for equal ra-

tios, a preference for smaller epochs and shorter slid-

ing window lengths.

Figure 6: Classiﬁer performance for control group (diago-

nal lines represent epoch-lagged conﬁgurations).

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

148

4 DISCUSSION

The goal of this study is to shed light on the disparity

of pipelines and classiﬁcation performances reported

in MI literature, by performing a systematic compar-

ative study of pipelines and their parameters on two

populations. This research is essential for the devel-

opment of MI-based neurorehabilitation systems as

processing pipelines and their speciﬁc conﬁgurations

yield optimal settings that can generalize from age-

matched participants to stroke patients. Nevertheless,

this study is also relevant for any MI paradigm be-

cause it reports on epoch-lagged conﬁgurations that

are suitable for applications other than neurorehabili-

tation. Therefore, to understand the behavior and rela-

tion of performance with different pipeline conﬁgura-

tions and age groups, this section discusses the corre-

lation between feature quantity and classiﬁcation per-

formance, accuracy performance between pipelines,

and the impact of age on performance.

4.1 Features and Classiﬁer Performance

The results reveal that classiﬁer performance de-

creases as epoch and sliding window lengths and their

within ratio increase, with signiﬁcant, negative corre-

lation test results (Table 1, Table 2). These inverse

correlations can be related to overﬁtting (Le

on et al.,

2020). The large quantity of features provides the

classiﬁer with too much detail and noise, allowing

the algorithm to overﬁt the training data. As part

of the training process for the classiﬁer, a portion of

the training data is withheld as new data for cross-

validation testing. Because of this, we are able to

see a negative impact on the classiﬁcation of unseen

data. This might explain why classiﬁer performance

decreases when feature quantity increases.

4.2 Pipeline Performance

Pipeline performance best results are summarized in

Table 3. Epoch-lagged conﬁgurations were tested and

yielded the highest performances. However, they are

inadequate for rehabilitative methods that rely on real-

time feedback because they add a time lag equiva-

lent to the epoch and window length. Online, real-

time feedback induces more pronounced attention and

motor cortical activation; it works as a hybrid con-

trol that augments sensory-feedback processing (Ros

et al., 2014), improving BCI performance when efﬁ-

ciently delivered (Ramos-Murguialday et al., 2019),

and eliminates the need for extensive sessions involv-

ing a high number of trials (> 100) (Ortner et al.,

2015). Therefore, it is desirable to use non-lagged

conﬁgurations with a window length shorter than the

epoch length so that the lag time is as short as the in-

terval length, meaning the classiﬁer output happens at

each sliding interval of the sliding window.

The 4-CSP-Hamming pipeline outperformed the 2-

CSP-Boxcar pipeline for all of the conﬁgurations in

both populations. This can be related to the number

of CSP output channels with a slightly better perfor-

mance seen for conﬁgurations with 4 CSP channels

versus 2 CSP channels (M

uller-Gerking et al., 1999).

Nevertheless, both pipeline performances are in the

adequate control range for MI-based BCIs (≥ 70%)

(Ortner et al., 2015).

4.3 Impact of Age on Performance

Age matched group obtained similar performance as

control group, with no signiﬁcant differences for all

tested conﬁgurations. Our results show that using

conﬁgurations with an epoch/window ratio of 2 ob-

tain similar results to the extended training with high

high number of trials (80 per session) and 6 sessions

during training, whose maximum performance was

80.7% and minimum was 72.4% (Ortner et al., 2015),

and better than 66% for older participants (72.0±8.07

years old) (Chen et al., 2019) with lower SMR later-

alization for MI according to aging processes. This is

promising for neurorehabilitation adapted for stroke

patients because shorter sessions might help alleviate

the exhaustiveness of MI training.

4.4 Limitations

This work had some limitations such as the low num-

ber of participants, equipment differences in both

EEG and VR rendering modalities between popula-

tions and gender unbalance. Moreover, the existence

of pipelines not used in this study, and the applicabil-

ity of the results for a stroke population. The number

of participants in this study, (N=12), might be too low

to achieve a high statistical power, but it is greater

than the number of participants recruited for BCI per-

formance studies that used datasets from BCI com-

petitions (N=4) and have helped to improve classiﬁer

performance. Inconsistencies between the EEG de-

vices and electrode placements might have enhanced

the differences between the two populations. Nev-

ertheless, the protocol used to process the data for

both groups of participants was the same. Impedance

quality was kept on the same level, identical pipelines

were used, and non-statistical differences were found

between groups. Although it might be interesting

to check with several classes of pipelines, doing a

larger comparison would be impracticable. The age-

Finding the Optimal Time Window for Increased Classiﬁcation Accuracy during Motor Imagery

149

Table 3: Sum-up of results.

MI-performance

Conﬁgurations Control Age-matched Diff. between-groups

Epoch-lagged

4-CSP-Hamming 90 85 5

2-CSP-Boxcar 74.25 76.67 -2.42

Diff. between-pipelines 15.75 8.33 X

Non Epoch-lagged

4-CSP-Hamming 79.68 79.09 0.59

2-CSP-Boxcar 72.41 71.25 1.16

Diff. between-pipelines 17.27 7.84 X

Diff. Lagged and non-lagged

4-CSP-Hamming 10.32 5.91 4.41

2-CSP-Boxcar 1.84 5.42 -3.58

matched group is a step closer, but stroke patients may

perform differently due to their speciﬁc lesion types.

5 CONCLUSION

Two populations were recruited for the investiga-

tion of age on MI classiﬁcation, and two CSP-based

pipelines were compared to ﬁnd the optimal pipeline

settings for improving classiﬁer performance. Classi-

ﬁcation performance under multiple pipeline settings

for MI in both groups showed no signiﬁcant effect for

age. The results conﬁrmed that the number of fea-

tures used to train a classiﬁer depends on the epoch-

window lengths relation. Classiﬁer performance in-

creased when a smaller epoch/window ratio was ap-

plied. However, using an epoch/window ratio equal

to one, introduce a lag in time equal to the length

of the epoch, which is not desirable for online pur-

poses. By investigating various time parameters that

directly affect classiﬁcation performance, an optimal

pipeline conﬁguration was found comprised of the 4-

CSP-Hamming pipeline using an epoch/window ratio

of 2 and ofﬂine of 0.1 seconds for both populations.

5.1 Future Work

New datasets will be acquired for the ongoing project,

including from stroke participants, increasing the

population size and the statistical power of the study.

We also plan to compare online classiﬁer performance

versus training classiﬁer performance closing the ac-

curacy gap and achieving a more generalizable classi-

ﬁer.

ACKNOWLEDGEMENTS

This work was supported by by the Portuguese

Foundation for Science and Technology through

the NeurAugVR FCT project (PTDC/CCI-

COM/31485/2017), the NOVA-LINCS-NOVA

Laboratory for Computer Science and Informatics

(PEest/UID/CEC/04516/2019), the Laborat

orio

de Rob

otica e Sistemas em Engenharia e Ci

encia

(UIDB/50009/2020) and Scientiﬁc Employment

Stimulus (CEECIND/01073/2018).

REFERENCES

Ahn, M., Cho, H., Ahn, S., and Jun, C. (2013). High

theta and low alpha powers may be indicative of bci-

illiteracy in motor imagery. PLoS One, 8(11):1303–

1309.

Ahn, M. and Jun, S. C. (2015). Performance variation in

motor imagery brain–computer interface: A brief re-

view. Journal of Neuroscience Methods, 243:103–

110. https://doi.org/10.1016/j.jneumeth.2015.01.033.

Ai, Q., Liu, Q., Meng, W., and Xie, S. Q. (2018).

Advanced Rehabilitative Techonology, chapter EEG-

Based Brain Intention Recognition, pages 135–166.

Academic Press. https://doi.org/10.1016/B978-0-12-

814597-5.00006-0.

Bai, Z., Fong, K. N. K., Zhang, J. J., Chan, J., and Ting,

K. H. (2020). Immediate and long-term effects of

bci- based rehabilitation of the upper extremity after

stroke: a systematic review and meta- analysis. Jour-

nal of NeuroEngineering and Rehabilitation, 2:1–20.

https://doi.org/10.1186/s12984-020-00686-2.

Cervera, M. A., Soekadar, S. R., Ushiba, J., del

R. Mill

an, J., Liu, M., Birbaumer, N., and Garipelli,

G. (2018). Brain-computer interfaces for post-

stroke motor rehabilitation: a meta-analysis. An-

nals of Clinical and Translational Neurology, 5(5).

https://doi.org/10.1002/acn3.544.

Chen, C., Chen, P., Belkacem, A. N., Lu, L., Xu, R., Tan,

W., Li, P., Gao, Q., Shin, D., Wange, C., and Ming,

D. (2020). Neural activities classiﬁcation of left and

right ﬁnger gestures during motor execution and mo-

tor imagery. Brain-Computer Interfaces, pages 1–11.

https://doi.org/10.1080/2326263X.2020.1782124.

Chen, M. L., Fu, D., Boger, J., and Jiang, N.

(2019). Age-related changes in vibro-tactile

eeg response and its implications in bci ap-

plications: A comparison between older and

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

150

younger populations. IEEE Transactions on Neu-

ral Systems and Rehabilitation Engineering, 27(4).

https://doi.org/10.1109/TNSRE.2019.2890968.

Irimia, D. C., Ortner, R., Poboroniuc, M. S., Ignat,

B. E., and Guger, C. (2018). High classiﬁca-

tion accuracy of a motor imagery based brain-

computer interface for stroke rehabilitation train-

ing. Frontier in Robotics and AI, 5(130):9.

https://doi.org/10.3389/frobt.2018.00130.

Khan, M. A., Das, R., Iversen, H. K., and Puthussery-

pady, S. (2020). Review on motor imagery

based bci systems for upper limb post-stroke

neurorehabilitation: From designing to applica-

tion. Computers in Biology and Medicine, 123.

https://doi.org/10.1016/j.compbiomed.2020.103843.

on, J., Escobar, J. J., Ortiz, A., Ortega, J., Gonz

alez, J.,

Mart

ın-Smith, P., Gan, J. Q., and Damas, M. (2020).

Deep learning for eeg-based motor imagery classiﬁ-

cation: Accuracy-cost trade-off. Plos One, 15(6).

https://doi.org/10.1371/journal.pone.0234178.

Lotte, F., Bougrain, L., Cichocki, A., Clerc, M.,

Congedo, M., Rakotomamonjy, A., and Yger, F.

(2018). A review of classiﬁcation algorithms for

eeg-based brain–computer interfaces: a 10 year up-

date. Journal of Neural Engineering, 15(3):28.

https://doi.org/10.1088/1741-2552/aab2f2.

uller-Gerking, J., Pfurtscheller, G., and Flyvbjerg,

H. (1999). Designing optimal spatial ﬁlters

for single-trial eeg classiﬁcation in a movement

task. Clincial Neurophysiology, 5(1):787–798.

https://doi.org/10.1016/S1388-2457(98)00038-8.

Mubin, O., Alnajjar, F., Mahmud, A. A., Jishtu, N., and

Alsinglawi, B. (2020). Exploring serious games for

stroke rehabilitation: a scoping review. Disability

and Rehabilitation: Assistive Technology, pages 1–7.

https://doi.org/10.1080/17483107.2020.1768309.

Oldﬁeld, R. C. (1971). The assessment and analysis of

handedness: The edinburgh inventory. Neuropsy-

chologia, 9(1):97–113. https://doi.org/10.1016/0028-

3932(71)90067-4.

Ortner, R., Scharinger, J., Lechner, A., and Guger, C.

(2015). How many people can control a motor im-

agery based bci using common spatial patterns? In

2015 7th International IEEE/EMBS Conference on

Neural Engineering (NER), pages 202–205, Montpel-

lier. IEEE. https://doi.org/10.1109/ner.2015.7146595.

Padﬁeld, N., Zabalza, J., Zhao, H., Masero, V., and Ren,

J. (2019). Eeg-based brain-computer interfaces using

motor-imagery: Techniques and challenges. Sensors,

19:1–34. https://doi.org/10.3390/s19061423.

Pfurtscheller, G. and Lopes, F. H. (1999). Event-related

eeg/meg synchronization and desynchronization: ba-

sic principles. Clinical Neurophysiology, 110:1842–

1857. https://doi.org/10.1016/S1388-2457(99)00141-

Pfurtscheller, G., Neuper, C., Obermaier, B., Krausz,

G., Scherer, R., Graimann, B., , and Schrank,

C. (2003). Graz-bci: state of the art and clin-

ical applications. IEEE Transactions on Neural

Systems and Rehabilitation Engineering, 11(2):1–4.

https://doi.org/10.1109/tnsre.2003.814454.

Putze, F. (2019). Brain Art: Brain-Computer Inter-

faces for Artistic Expression, chapter Methods and

Tools for Using BCI with Augmented and Vir-

tual Reality, pages 433–446. Springer, Cham.

https://doi.org/10.1007/978-3-030-14323-7.

Ramos-Murguialday, A., Curado, M. R., Broetz, D., Yil-

maz, ., Brasil, F. L., Liberati, and G., . . . Birbaumer,

N. (2019). Brain-machine interface in chronic

stroke: Randomized trial long-term follow-up. Amer-

ican Society of Neurorehabilitation, 33(3):188–198.

https://doi.org/10.1177/1545968319827573.

Renard, Y., Lotte, F., Gilbert, G., Congedo, M., Maby,

E., Delannoy, V., Bertrand, O., and L

ecuyer, A.

(2010). Openvibe: An open-source software plat-

form to design, test and use brain-computer inter-

faces in real and virtual environments. Presence

Teleoperators and Virtual Environments, 19(1):35–53.

https://doi.org/10.1162/pres.19.1.35.

Ros, T., Baars, B. J., Lanius, R. A., and P.Vuilleumier

(2014). Tuning pathological brain oscillations

with neurofeedback: a systems neuroscience frame-

work. Frontiers in Human Neuroscience, 8:1008.

https://doi.org/10.3389/fnhum.2014.01008.

Suryotrisongko, H. and Samopa, F. (2015). Eval-

uating openbci spiderclaw v1 headwear’s

electrodes placements for brain-computer

interface (bci) motor imagery application.

Procedia Computer Science, 72:398–405.

https://doi.org/10.1016/j.procs.2015.12.155.

Vourvopoulos, A. (2018). Using Brain-Computer Interac-

tion and Multimodal Virtual-Reality for Augmenting

Stroke Neurorehabilitation. PhD thesis, Universidade

da Madeira. https://doi.org/10400.13/1992.

Vourvopoulos, A., Ferreira, A., and Bermudez, S.

(2016). Neurow: An immersive vr environment

for motor-imagery training with the use of brain-

computer interfaces and vibrotactile feedback. In

3rd International Conference on Physiological Com-

puting Systems, Lisbon, Portugal. SCITEPRESS.

https://doi.org/10.5220/0005939400430053.

Vourvopoulos, A., Jorge, C., Abreu, R., Figueiredo, P.,

Fernandes, J., and Bermudez, S. (2019). Efﬁ-

cacy and brain imaging correlates of an immer-

sive motor imagery bci-driven vr system for upper

limb motor rehabilitation: A clinical case report.

Frontiers in Human Neuroscience, 13(244):567–577.

https://doi.org/10.3389/fnhum.2019.00244.

Wang, J., Feng, Z., Ren, X., Lu, N., Luo, J., and Sun,

L. (2020). Feature subset and time segment selec-

tion for the classiﬁcation of eeg data based motor im-

agery. Biomedical Signal Processing and Control, 61.

https://doi.org/10.1016/j.bspc.2020.102026.

Finding the Optimal Time Window for Increased Classiﬁcation Accuracy during Motor Imagery

151