IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

71, 2022 4010510

CCA-Based Compressive Sensing for SSVEP-Based

Brain-Computer Interfaces to Command a
Robotic Wheelchair
Hamilton Rivera-Flor , Dharmendra Gurve , Member, IEEE, Alan Floriano , Denis Delisle-Rodriguez ,
Ricardo Mello , and Teodiano Bastos-Filho

Abstract— People with severe physical disabilities are not able activities, being robotic wheelchairs widely used to aid their
of using standard robotic wheelchairs, which generally demand mobility [1]. Most of these devices require the user’s motor
some motor skills, and therefore total usage of associate muscles. ability to manipulate a joystick or keyboard for conveying con-
Robotic wheelchairs commanded by brain-computer interfaces
(BCIs) based on electroencephalography have demonstrated to trol commands during locomotion [1], [2]. However, people
be an alternative for these end-users. In general, existing robotic suffering severe motor impairments due to neural diseases,
wheelchairs commanded by BCIs require special platforms such as amyotrophic lateral sclerosis (ALS) and locked-in
adapted to the EEG-BCI, and end-users need to attend a long syndrome (LIS) are unable of using robotic wheelchairs
training process to safely drive these devices. But many times demanding their total muscle usage [1]. To overcome this
these potential users do not have access to training sessions;
due to mobility problems or technology access restrictions. issue, systems operated with few or without residual mus-
This study proposes an EEG-based BCI with a customizable cular movements have been proposed [1], [3]. For instance,
configuration to be used in cloud architectures for the remote eye blinks, eye movements, facial/head gestures, blowing
control of robotic wheelchairs. This research explores two types and sucking (also named “sip and puff switch”), and brain
of steady-state visual evoked potential (SSVEP)-based BCI by waves [2] have successfully been used to command these
applying canonical correlation analysis (CCA) and compressive
sensing (CS) as a novelty, adopting one free calibration, and the robotic systems. In particular, brain wave-based systems, such
other including a calibration stage. The free-calibrated SSVEP as brain–computer interfaces (BCIs), have been developed by
recognition approach (CS-ncCCA) using compression ratio (CR) many researchers to acquire and translate brain signals into
at 60% obtained accuracy (ACC) of 85% and information control actions [2], [3], [4], [5], [6], [7], [8], [9], [10]. BCIs
transfer rate (ITR) of 102 bits per minute (b/min), whereas the are systems that allow the control of external devices by using
calibrated BCI (CS-wcCCA) applying also CR at 62% achieved
ACC of 85% and ITR of 195 b/min. As a highlight, the proposed the brain’s signals. Because of BCIs do not use neuromuscular
BCI allows a significant reduction of the transmitted file size commands as input, the advent of these interfaces allows
(TFS) and improves the communication latency that may be to establish an alternative pathway for people with severe
useful in remote and cloud robotics applications, such as for motor impairment to interact with the world. The electroen-
users with severe motor disabilities, to train driving safely robotic cephalogram (EEG) becomes the most widely used technique
wheelchairs via Internet of Things (IoT).
for brain signal acquisition in BCI systems, due to its high
Index Terms— Brain computer interfaces (BCIs), compressive temporal resolution, noninvasiveness, relatively low cost, and
sensing (CS), robotic wheelchairs, steady-state visual evoked portability [4], [11], [12].
potentials (SSVEPs).
Nowadays, assistive BCI technologies based on steady-state
visual evoked potentials (SSVEPs) responses using EEG are
I. I NTRODUCTION
being widely applied for mobility, communication, and enter-

F OR people with severe physical disabilities is very impor-

tant to have independence for performing daily living
tainment [4], [6], [11], [12], [13], due to its robustness for
target recognition with few or absence of calibration steps,
Manuscript received 11 March 2022; revised 26 September 2022; increasing usability [4], [6], [14]. The SSVEP responses can
accepted 9 October 2022. Date of publication 31 October 2022; date of current be elicited in the visual cortex after stimulating the user’s
version 15 November 2022. This work was supported in part by the Federal eyes with light stimuli flickering at a constant frequency [6],
University of Espírito Santo, in part by the Conselho Nacional de Desenvolvi-
mento Científico e Tecnológico (CNPq), and in part by the Coordenação de [8], [15]. As a result, these potentials are manifested on the
Aperfeiçoamento de Pessoal de Nível Superior (CAPES)/Brazil under Grant EEG signal as an oscillatory component at the same frequency
88887.095636/2015-01. The Associate Editor coordinating the review process and/or harmonics of the visual stimulus [3], [4], [8], [16].
was Chi-Hung Hwang. (Corresponding author: Hamilton Rivera-Flor.)
Hamilton Rivera-Flor, Teodiano Bastos-Filho, and Ricardo Mello are with Then, visual stimuli oscillating for specific frequencies can
the Department of Electrical Engineering, Federal University of Espírito be codified through a BCI to convey different commands [4],
Santo, Vitória 29075-010, Brazil (e-mail: [email protected]). [6], [13]. For instance, an SSVEP-based BCI system to control
Denis Delisle-Rodriguez is with the Edmond and Lily Safra International
Institute of Neurosciences, Macaíba 59288-899, Brazil. electric wheelchairs is proposed in [12] to assist individuals
Dharmendra Gurve is with the Department of Electrical Engineering, suffering from neuromuscular degenerative diseases, such as
Ryerson University, Toronto, ON M5B-2K3 Canada. ALS and LIS. In [1] and [2], other BCI based on SSVEP to
Alan Floriano is with the Institute of Mathematics and Computer Sciences,
University of São Paulo, São Carlos 13566-590, Brazil. command an autonomous robotic wheelchair is also proposed,
Digital Object Identifier 10.1109/TIM.2022.3218102 in which four targets are decoded by using EEG feature
1557-9662 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 4010510 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

extraction and a rule-based classifier. Teleoperated robotic times these potential users do not have access to training
systems through augmentative reality (AR) scenarios are sessions, due to their mobility problems, access restrictions
becoming a potential tool in BCIs for training novice users of to laboratories, nonsuitable infrastructure for people with
wheelchairs [17], as well as to improve physical rehabilitation disabilities, and a lack of support for training, among others.
outcomes [18], with a wide variety of benefits including safe We hypothesize that an EEG SSVEP-based BCI may facilitate
controlled environments, low-cost, and flexibility [17], [19]. the safe training of novice users of wheelchairs at home
Regarding robotic wheelchairs controlled by SSVEP-based via IoT.
BCIs, teleoperation through AR may tackle some issues and The main topics addressed in this current research are:
challenges, such as mobility restrictions because of COVID-19 robotic wheelchairs training solutions for people with severe
pandemic [20], as well as financial or health limitations of disabilities [17], [18], [19]; SSVEP-based BCI system for
people with severe physical impairments. Fortunately, recent commanding robotic devices [7], [29], [30], [31], [32],
advances in BCI, Internet of Things (IoT) [21], and cloud [33], [34]; and CS in biomedical signals, IoT, and remote
computing can make it possible to build teleoperated systems, health [21], [22], [23], [24], [26].
such as robotic wheelchairs in virtual environments that allow This work aims to develop an SSVEP-based BCI using
novice persons to remotely learn the control commands of CS on EEG signals, in order to increase the ITR when
these devices. commanding a robotic wheelchair remotely via IoT. The
A BCI based on wearable EEG headsets and cloud proposed system reduces the size of data to be transferred,
computing to detect and recognize the user’s mental also reducing computational cost without losing signal quality,
states is proposed in [21]. The architecture of this system and obtaining high accuracy and ITR. The proposed system
acquires EEG signals and performs lightweight on-site signal worked appropriately in a virtual environment or simulator,
preprocessing tasks to afterward transmit the EEG data to a and the future will be tested in Cloud-BCI architecture.
cloud service environment, which runs computation-intensive
signal processing. Compressive sensing (CS)-based signal II. M ATERIALS AND M ETHODS
compression and sensing have gained a wider acceptance, as a Fig. 1 shows a flowchart of an SSVEP-based BCI robotic
sampling technique for sampling signals at their information wheelchair for training remotely novice users. Then, people
rate. CS takes the advantage of sparsity or compressibility of with severe motor disabilities can remotely command a wheel-
the underlying signal to simultaneously sample and compress chair by using their brain signals. As a novelty, this approach
the signal. CS has applications in biomedical fields such applies, after EEG acquisition, a compression method based on
as biomedical imaging, and it also has been applied for CS to reduce the data size but preserving the signal’s quality,
processing biological signals, such as electrocardiogram and increasing the ITR. As another contribution, the proposed
(ECG), electroencephalographic, and other physiological system is designed to remotely process the compressed sig-
signals by exploiting the sparsity present in their features. nals without reconstruction, greatly enhancing the processing
The various Wireless Sensor Networks in which CS has been speed. Alternatively, our system contains parameters easily
applied are wireless body area networks [22], brain-machine configurable to increase the performance of each user. As an
interface [23], wireless surface electromyography (EMG) [24] advantage, the wheelchair control system is implemented in
for telehealth monitoring, and EEG compression [25], [26]. the platform robot operating system (ROS) [35], which stan-
CS has a strong mathematical foundation, but some challenges dardizes the communication, connection, and control between
are still addressed in the field. For instance, to obtain a simple, different wheelchairs, simulators, or robots. Details about
efficient, and universal CS acquisition strategy that can be the proposed system, dataset, and experiment conducted for
applied to diverse signals, also guaranteeing fast data acquisi- evaluation are presented below.
tion. Similarly, the development of a universal CS reconstruc-
tion algorithm, faster, robust, less complex, and guaranteeing A. Wheelchair Training Platform—Overview
convergence is a research scopes still opened [27]. CS can The proposed SSVEP-based BCI is composed of three
be used in cloud-BCI applications to facilitate high-speed modules: 1) data acquisition; 2) BCI based on SSVEP-CS;
signal processing, saving also memory space and time during and 3) control module for the robotic wheelchair. The first
signal acquisition and transmission. Then, the CS framework module for EEG acquisition and preprocessing is divided into
for EEG processing can help to tackle current main problems three main steps: visual stimulation; EEG signal recording;
and challenges that still limit the use of cloud-BCI systems EEG filtering and compression by applying CS at the sensor
for teleoperation applications, such as transmission of large node after conveying EEG data. After, the SSVEP-based BCI
amounts of EEG data, fast processing, high performance module receives the compressed data to translate it into control
and scalability, and energy-saving for long-term battery commands. Finally, the third module for control is composed
operation [28]. Furthermore, the development of CS-based of a virtual or real robotic wheelchair. The virtual wheelchair
BCIs for achieving both high accuracy (ACC) and information (the Simcadrom) simulates an electric-powered wheelchair
transfer rate (ITR) is desirable for practical application. (EPW) in a test room for training purposes and testing
In general, existing robotic wheelchairs commanded control strategies [17]. The real wheelchair is equipped with
by BCIs require powerful hardware for high-speed signal a liquid-crystal display (LCD) screen that also alternatively
processing and a special wheelchair adapted to the EEG-BCI. displays an environment of navigation providing a menu of
In addition to these requirements, end-users need to attend a options for local configuration [17]. Then, users can select
long training process to safely drive these devices. But many remotely the desired command (or direction) to move or turn
licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 RIVERA-FLOR et al.: CCA-BASED CS FOR SSVEP-BASED BCIs TO COMMAND A ROBOTIC WHEELCHAIR 4010510

a frequency range from 7 to 50 Hz to finally apply CS and

transmit the compressed data.
CS, also known as compressive sampling, is an emerging
data acquisition technique [28]. Specifically, for sparse signals,
CS combines compression and sampling together for reducing
considerably the sampling rate. The advantage of combining
compression and sampling in one stage has been demonstrated
for various applications in previous studies [25], [28].
As a highlight, a reconstruction-free CS for SSVEP recogni-
tion directly using compressed measurement is proposed here.
Let x ∈ R E×t be an EEG epoch corresponding to an SSVEP
stimulus, where E and t are the total number of electrodes
and the epoch length, respectively. Then, applying CS to a
set of EEG epochs X ∈ R E×N where N is the current deep
learning (DL), compressed EEG epochs with the desirable new
DL S (or measurement samples) are obtained, being N > S.
In short, a random matrix A ∈ R S×N is first generated to
immediately calculate its orthonormal matrix  ∈ R S×N , also
termed the measurement matrix. Afterward, this measurement
matrix is used in our proposed schemes using CS without a
signal reconstruction step by using (2)
 T
 = orth AT (1)
Y = X (2)

where ( . )T is the transpose of a matrix, orth( . ) is the

orthonormal vector calculation, and Y is the compressive
measurement of size S. It is worth noting that  is used in
both sensor node and BCI.
After transmitting the compressed data an SSVEP classifier
can be used, including a calibration stage or without cali-
bration. For calibration-free or not calibrated SSVEP clas-
sification, it is combined CCA [36] with CS (this approach
termed here as CS-ncCCA), whereas the proposed method
based on Nakanishi et al. [37] with calibration is used on the
compressed EEG data (approach termed here as CS-wcCCA).
2) CCA-Based CS: CCA is a multidimensional statistical
Fig. 1. Block diagram of the proposed BCI based on CS and SSVEP. analysis technique that finds underlying linear correlations
the wheelchair. The module receives commands through a mini between a set of multichannel EEG data and another set of
PC Intel NUC attached at the back of the wheelchair, which reference signals [36]. This reference set consists of sine and
allows the remote control by SSVEP. cosine signals at the fundamental and harmonic frequencies of
each stimulus. In our case, both sets with EEG and reference
B. Proposed BCI Using CS signals are transformed by applying (1), using the matrix .
A configurable system is proposed here to select different As a result, the canonical correlation indices are calculated for
window sizes or data lengths (DLs) for processing, different each reference signal (our visual stimulus frequency), choosing
levels of data compression, and different classification tech- as classification outputs the stimulus that presents the highest
niques by developing an algorithm that can be customized correlation value.
and optimized for each user, increasing the ITR. 3) SSVEP Classification With Calibration: This approach
This study explores two types of SSVEP-based BCI: is an extension of CCA [37], and incorporates as a novelty a
1) calibration-free by combining as a novelty the canonical training stage with subject-specific compressed EEG data, aim-
correlation analysis (CCA) [36] with a compression matrix to ing to improve the BCI accuracy. Training reference SSVEP
transform synthesized sine and cosine waveforms that oscillate signals are obtained by averaging multiple SSVEP trials. Cor-
at the frequency of visual stimuli and 2) with calibration stage relation coefficients between projected EEG epochs (from the
by using on compressed EEG data the method proposed by testing set) and synthesized reference signals based on sine and
Nakanishi et al. [37]. cosine waveforms at the fundamental and harmonic frequen-
1) Sensor Node Processing: For both BCIs at the sensor cies, representing all visual stimuli (or targets) are obtained.
node, the EEG signals are first referenced by applying com- Then, the classification output is associated with a target
mon average reference (CAR), and after bandpass filtered in with a maximum correlation coefficient. As the calibration

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 4010510 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

stage uses the compressed EEG data, the classification model

can be extended to new compressed data without additional
transformation, reducing the computational costs.

C. Dataset
The international SSVEP benchmark dataset [37], [38] was
employed here to evaluate both proposed BCI schemes, which
gathers EEG recordings from 35 healthy subjects (17 females,
aged 17–34 years, mean age: 22 years) who focused on
40 characters flickering at different frequencies (from 8.0 to
15.8 Hz with an interval of 0.2 Hz). For each subject, the
experiment consisted of six blocks. Each block contained
40 trials corresponding to 40 characters displayed randomly.
Each cue appeared on the screen for a period of 0.5 s indicating
to each subject the target, and consequently, she/he shifted
Fig. 2. Block diagram of the experimental setup scenario.
her/his gaze as soon as possible within the cue duration. After
disappearing the cue, all stimuli started to flicker on the screen
In this study a quantitative test of the CS performance is
for a period of 5 s, followed by a blank screen for a period
carried out by analyzing a set of prerecorded EEG data, where
of 0.5 s before beginning the next trial.
N is a data vector sample frame of background scalp EEG,
EEG data through a 64 electrodes-cap aligned according
this is based upon a Gaussian random measurement matrix
to the international 10–20 system were acquired with the
and the matching pursuit orthogonal reconstruction method
reference electrode on the vertex, using the Synamps2 system
as the number of measurement samples (S) is varied. This
(Neuroscan, Inc.) with a sampling rate at 1000 Hz, and
is equivalent to changing the CR, which corresponds to the
bandpass filter in a frequency range from 0.15 to 200 Hz.
fraction of the full data that is required to represent the signal.
Nine electrodes over the parietal and occipital areas (Pz, PO5,
CR can be represented mathematically by using the following
PO3, POz, PO4, PO6, O1, Oz, and O2) were used in our study.
equation:
D. Experimental Setup Scenario S
CR = × 100. (3)
Fig. 2 shows the experimental setup of our simulated N
evaluation by employing the benchmark dataset [37], [38]. The ITR in bits/min is calculated as follows:
 
In this dataset, the subjects focused on 40 characters flickering (1 − P)
Po = log2 (4)
at different frequencies. From these characters or stimuli, (M − 1)
a total of four groups composed of ten characters each were   60
created, which were associated in our simulation with four ITR = log2 M + P × log2 P + (1 − P) × Po × (5)
T
control signals (such as reverse, right, left, and forward),
where M is the total number of classes (40 stimuli in our case),
in order to convey commands to a virtual wheelchair during
P is the classification accuracy, and T (seconds/symbol) is the
navigation. Here, each control command was designed in a
time needed to deliver each command [15], [39].
frequency range as follows: reverse (any stimulus from 8.0 to
TFS is the size in bytes of the data packet to be transmitted
9.8 Hz), right (any stimulus from 10.0 to 11.8 Hz), left (any
over the network from the data acquisition module to the BCI
stimulus from 12.0 to 13.8 Hz), and forward (any stimulus
module, and it can be calculated by (6). TCP is the sum of
from 14.0 to 15.8 Hz). Then, we apply CS and transmit the
the processing time in the data acquisition module TDA and
compressed data using a ROS node. The remote BCI module
the processing time in the BCI module TBCI without taking
receives the compressed data to processing using the method
into account the transmission time
selected (CS-ncCCA or CS-wcCCA) to classify the control
commands and conveys these commands to the wheelchair TCP = TDA + TBCI . (6)
simulation system [17] through a ROS node. The command
received by the control module is sent through the node to the The latency is the estimated time or delay for data to travel
wheelchair simulator developed using Unity SimCadRom [17]. from its source to a destination across the network, and it is
It is worth mentioning that this system allows to convey control computed by (7). The latency estimation is a difficult process
commands to a virtual or real wheelchair. since it depends on the network and several variable factors,
such as communication protocols and bandwidth
E. Evaluation TFS(1 + %OD)
Latency = (7)
To evaluate the performance of the proposed methods by NDR
using different setups with different DL and compression where %OD is the extra bytes of format information that each
ratio (CR), five metrics were used; such as ACC and ITR, packet requires. That additional data reduces the effective rate
transmitted file size (TFS), time-consuming for processing of transmission but provides the benefit of error-free transfers
(TCP), and latency [25], [37], [39], [40], [41]. (typical range from 10% to 30% of the original data size).

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 RIVERA-FLOR et al.: CCA-BASED CS FOR SSVEP-BASED BCIs TO COMMAND A ROBOTIC WHEELCHAIR 4010510

Fig. 3(c) shows the relationship between ACC and ITR for
different DLs and CR, applying CS-ncCCA. The CS-ncCCA
approach using 4 s-EEG epochs reached the highest and
most consistent accuracy with moderated ITR, when compared
with other setups. For this DL of 4 s, the accuracy and
ITR decreased less than 4% for different CR from 0% to
80%. As another result, an ITR higher than 100 b/min,
and acceptable ACC (around 80%) were achieved for CR
up to 60%, being an optimal performance when considering
previous works [38], [41].
These findings are encouraging for future practical imple-
mentation in cloud-BCIs, as less EEG data at the sensor
node can be transmitted via internet, but performing well the
SSVEP recognition. However, this calibration-free approach
did not perform well when using the shortest DLs (1 and
0.5 s), decreasing significantly its accuracy (ACC < 45%).
We observed that the best ITR was obtained by employing
DLs of 2 and 1 s, confirming the relevance of using CS in a
cloud BCI.
Furthermore, we also noted that applying CS plus a calibra-
tion step (CS-wcCCA scheme) increased both ACC and ITR,
as shown in Fig. 3(d) and (e). As a result, this approach using
window sizes of 4 s attained the highest ACC, being 95.89% ±
10.77% for CR of 0%, and 94.77% ± 13.67% for CR of 80%.
In contrast for this same window size, it decreased consider-
Fig. 3. Performance reached with the recognition system by combining CS
and a calibration-free stage-based CCA (termed as CS-ncCCA) or a calibration
ably the ITR up to 72.12 ± 11.77 b/min for CR of 0%, and
stage-based CCA (termed as CS-wcCCA). (a) ACC versus compression rate 71.58 ± 14.47 b/min for CR of 80%. We observed again that
(CR); (b) ITR versus CR; (c) ACC versus ITR; (d) ACC versus CR; (e) ITR the accuracy diminished when decreasing the window sizes
versus CR; (f) ACC versus ITR.
for EEG processing. For instance, CS-wcCCA achieved ACC
around 84% and 81.37% for DLs of 2 and 1 s, respectively.
NDR is the network data rate which varies for each technology. However, the highest ITR (around 221.86 b/min for CR of 0%,
For instance, NDR varies from 150 to 450 Mb/s for 4G and 144.86 b/min for CR of 80%) was accomplished when
technology. The computer used for processing was an AMD using 1 s window length for EEG processing, instead of 2 s
Ryzen 5 3500U, RAM 8 GB, and processor 64×. The ANOVA (ITR < 130 b/min). It is worth noting that the CS-wcCCA
method (α = 0.05) was used on ACC and ITR values to obtain scheme using 1 s window length achieved the best perfor-
the best setup (DL and CR) when applying both CS-ncCCA mance (ACC and ITR of 80% and 200 b/min, respectively)
and CS-wcCCA approaches. for CR between 0% and 60%. Also, note that the approach
using the shortest DL of 0.5 s obtained ACC < 50% with
III. R ESULTS high ITR. These findings demonstrate the relevance of using
Both CS-ncCCA and CS-wcCCA approaches were evalu- a calibrated BCI to enhance both the SSVEP recognition and
ated with different compression rates (CR equal to 0%, 20%, ITR, as shown in [37] and [38]. As a highlight, a comparable
40%, 60%, and 80%), window sizes or DL (DL equal to 0.5, performance can be obtained by applying on 1 s EEG epochs
1, 2, and 4 s) and overlapping of 100 ms for EEG processing. the combination of CS with CR up to 60% and CCA with
Fig. 3(a) and (b) show that CS-ncCCA, when using DL of 4 s, subject-specific calibration. In general, the best ACC values
obtained the highest accuracy (ACC of 93.53% ± 10.04%) for for different CR were reached on DLs of 4 and 2 s, but
CR equal to 0%, which after decreased slightly to 89.77% ± attaining relatively low ITR. In contrast, better ITR results
14.96% when increasing CR to 80%. However, this DL of were achieved for CR from 0% to 60% when using window
4 s did not affect positively the ITR. Note that a low ITR of sizes of 1 and 0.5 s.
70.63 ± 12.36 b/min and 66.58 ± 16.66 b/min was achieved Fig. 3(f) shows ACC values above 80% when applying
when applying CR of 0% and 80%, respectively. As expected, the CS-wcCCA method with window lengths from 1 to
the proposed CS-ncCCA scheme, when using window sizes 4 s, being the highest accuracy for 4 and 2 s for different
of 2 s, enhanced the ITR attaining values between 111.65 ± CR from 0% to 80%, but achieving low and moderated
41.26 b/min (for CR equal to 0%) and 79.96 ± 42.30 b/min ITR, respectively. It is worth mentioning that an equivalent
(for CR equal to 80%), decreasing its accuracy between performance was obtained when applying CS-ncCCA and
80.37% ± 20.63% and 64.11% ± 24.43%, respectively. It is CS-wcCCA on 4 s EEG epochs. Therefore, the CS-ncCCA
worth noting that a similar performance (ACC around 80%, approach can be more convenient of using with long DLs
and ITR nearest to 111 b/min) was obtained for CR from 0% because of its low-computational cost and implementation
to 60%, using CS-ncCCA with DLs of 2 s. simplicity. As a highlight, the recognition system based on

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 4010510 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

TABLE I TABLE III

ANOVA C OMPARISON R ESULTS FOR D IFFERENT DL U SING B OTH ITR, TFS, AND TCP C OMPARISON P ERFORMANCE FOR D IFFERENT
CS- NC CCA AND CS- WC CCA A PPROACHES . T HE S ETUP OF B EST S ETUP A PPROACHES
P ERFORMANCE CS- NC CCA (DL = 2 s) AND CS- WC CCA (DL = 1 s)
W ERE U SED AS R EFERENCE

TABLE II
ANOVA C OMPARISON R ESULTS FOR D IFFERENT CR U SING B OTH
order to increase the ITR via internet, but preserving the
CS- NC CCA AND CS- WC CCA A PPROACHES . T HE S ETUP OF B EST
SSVEP classification accuracy.
P ERFORMANCE CS- NC CCA (DL = 2 s, CR = 0%), AND CS- WC CCA
After confirming through the statistical analysis the CR
(DL = 1 s, CR = 0%) W ERE U SED AS R EFERENCE
and DL that improve the ITR without a significant ACC
decrease are used to configure the system and it reaches the
best performance, increasing ITR, TFS and TCP. Table III
shows the performance for different configurations. The first
and second rows in this table present a configuration for
ncCCA and wcCCA that reached the maximum ACC (94%
and 96%, respectively), but with the lowest results of ITR
(70 and 72 b/min). Moreover, the configuration showed in
CS-wcCCA enhanced considerably the accuracy and ITR over the third and fourth rows for ncCCA and wcCCA reached
short window lengths from 1 to 2 s, especially when using the highest ITR (114 and 222 b/min respectively) but the
CR up to 60%. As another finding, we observed a consistent ACC decreased (80% and 82%). Furthermore, in the fifth and
performance (few ACC and ITR variations) when using the sixth rows for ncCCA and wcCCA the ACC and ITR were
CS-wcCCA approach on EEG epochs with a window length maximized, but TFS increased (10 678 and 5339 kB). In order
bigger than 2 s. Also, the results reached for 1 s window to obtain the optimal configuration parameters, we calculated
lengths with CR up to 40% are encouraging and consistent, the ITRmax to determine the best DL, as shown in Fig. 3(c)
obtaining ITR of 210.71 ± 78.57 b/min and acceptable ACC and (f). Afterward, the CR that did not produce a significant
of 78.94% ± 21.87% for practical implementation, as demon- difference for ACC and ITR was determined through ANOVA
strated in previous related studies [37]. analysis, as shown in Tables I and II. Then, the optimal
Notice that a high ITR is preferred in our proposed system. parameters were calculated for each subject, obtaining finally
Fig. 3 shows that the best ITR with good ACC was obtained the average performance for each tested setup. As a result,
by applying CR of 0% with CS-ncCCA (using DLs of 2 s) we observed that CS-ncCCA with CR at 60% achieved a
and CS-wcCCA (using DL of 1 s), respectively. Then, the per- smaller TFS 4271 kB, whereas CS-wcCCA with CR at 62%
formance of both approaches was compared through ANOVA obtained a smaller TFS (4271 and 2029 kB). These results
analysis (α = 0.05) with other setups that applied different show the potential of using CS to reach high ITR, and decrease
window lengths and equal CR of 0%. Table I shows the both the TCP and the TFS. Therefore, both CS-ncCCA and
statistical significance results, confirming that both CS-ncCCA CS-wcCCA approaches could improve the performance of a
and CS-wcCCA approaches using CR of 0% when using cloud-integrated BCI system.
2 and 1 s EEG epochs, respectively, enhanced significantly Table IV shows the latency obtained by simulating different
( p < 0.05) the SSVEP classification and ITR with respect generations of networks 3G (NDR from 384 kb/s to 2 Mb/s),
to other setups using different EEG epoch lengths. According 4G (NDR from 150 to 450 Mb/s), and 5G (over 10 Gb/s). The
to previous studies, these results show that our approaches latency was simulated for the TFS results of each method, with
applied in a cloud-BCI can increase its accuracy for a given an estimated OD of 20%. We observed that systems using CS
EEG epoch size (from 4 to 1 s), affecting also the ITR. Thus, reduced the latency two times and up to four times, compared
the system can be configured according to its performance on to systems without CS. These findings have significance for
each subject. future cloud robotic applications based on BCIs.
Table II also shows the results using ANOVA analysis The proposed system was tested in the simulation environ-
(α = 0.05) for different CR. We did not observe any ment SimCadRom [17] by considering the experimental setup
significant difference ( p > 0.05) of ACC and ITR values previously described in Section II-D. Our simulation consisted
when applying both CS-ncCCA and CS-wcCCA approaches of a virtual wheelchair commanded through the proposed
with CR from 0% to 60%. These findings suggest that the remote BCI by following a path, starting at point E and contin-
proposed approaches in a cloud-BCI system can be configured uing sequentially the points A-B-C-D-E, as shown in Fig. 4(a).
for each subject to transmit compressed data up to 60% in For this simulation, five subjects (S03, S07, S14, S22, and S29)

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 RIVERA-FLOR et al.: CCA-BASED CS FOR SSVEP-BASED BCIs TO COMMAND A ROBOTIC WHEELCHAIR 4010510

TABLE IV
S IMULATION OF L ATENCY FOR D IFFERENT G ENERATIONS OF N ETWORKS
FOR D IFFERENT S ETUP A PPROACHES

from the dataset (who presented the best performance in our

study) were selected, and their EEG data were used for decod-
ing the following four commands: reverse (from 8.0 to 9.8 Hz);
right (from 10.0 to 11.8 Hz); left (from 12.0 to 13.8 Hz);
and forward (from 14.0 to 15.8 Hz). Here, the CS-wcCCA
scheme using window lengths of 1 s, and CR at 60% was
used. It is worth noting that each control command was defined
for a group of ten consecutive frequencies. Fig. 4(b) shows
the adopted path reference with a black line, together with
other five paths followed independently by Subjects S03 (red
line), S07 (magenta line), S14 (green line), S22 (blue line),
and S29 (cyan line). As a result, the proposed CS-wcCCA
scheme reached a mean ACC of 87% ± 9.5%, averaged root
mean square error of 0.34 ± 0.047 m, and averaged time
of 192.1 ± 51.2 s for completing the reference path. This
performance demonstrates the feasibility of using the proposed
scheme for users to safely command a virtual wheelchair, and
also opens the possibility of future implementation with real Fig. 4. Simulation on SimCadRom by using the proposed CS-wcCCA scheme
with window lengths of 1 s and compression rate at 60%. (a) Scenario for
robotic wheelchairs. the wheelchair simulation on SimCadRom. (b) Adopted path (reference) with
black line, together with other five paths followed independently by subjects
S03 (red line), S07 (magenta line), S14 (green line), S22 (blue line), and S29
IV. D ISCUSSION (cyan line).
The proposed BCI system combining CS and CCA meth-
ods for EEG-SSVEP recognition reached good performance results suggest that the CS recovery for SSVEP recognition
when applying both calibration-free and calibration strategies. may be avoided to increase consequently the ITR, which is
In Table V a comparative summary of different approaches is desirable in our proposed work.
presented. We observed that the compressed data up to 60% We found that short window lengths of 1 s for EEG
can still provide enough information for target classification processing can produce high performance when a calibrated
without CS recovery at the receiver, reducing the locally com- system based on CCA and CS is employed instead of the
putational cost. These findings agree with previous research CCA calibration-free. As a result, the ITR enhances consid-
that applied CS on EEG signals [25], [26]. For instance, erably. The potential of CCA based on a training stage to
in [25], for an independent SSVEP-BCI based on CS, the increase the recognition accuracy (ACC ≥ 90%) and speller
results showed that for short DLs of 1 s and CR of 75%, ACC speed (ITR ≥ 150 b/min) was reported in [37] and [43].
of 72% was achieved, whereas ACC of 85% was obtained Chen et al. [43] developed a new joint frequency-phase mod-
for DLs of 4 s and CR of 75%. Other work [26] concluded ulation method and a user-specific SSVEP decoding algorithm
that when applying CR < 60% in an EEG SSVEP-based by using a filter bank CCA (FBCCA) method, obtaining
BCI using spatiotemporal sparse variational Bayes (STSVB) ACC and ITR of 88.92% and 259.2 b/min, respectively. Chen
the SSVEP detection accuracy became almost comparable et al. [38] used the FBCCA method in a calibration-free BCI,
with respect to uncompressed SSVEP data recognition. They enhancing significantly (averaged ACC of 91.95% and ITR of
also demonstrated with some experiments that their proposed 151.18 b/min) the standard CCA method. A similar frequency-
system continues performing well even at CR = 80%. Table V phase modulation strategy without applying FBCCA was
shows that our two approaches attained accuracy comparable previously adopted in [37], getting also good performance.
with respect to these two studies [25], [26], which applied CS Nakanishi et al. [29] used a task-related component analysis
recovery at the receiver. Also notice that the window length (TRCA), obtaining ACC and ITR of 89.8% and 325.3 b/min,
for EEG processing was reduced considerably by using the respectively, in offline tests, and 198.6 b/min in online tests.
CS-wcCCA method, in contrast to STSVB based on Laplace Table V also shows that various research proposed
(STSVB-L) [26], which also includes a learning stage. These SSVEP-based BCIs to command robotic systems, such as

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 4010510 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

TABLE V
C OMPARATIVE S UMMARY OF D IFFERENT A PPROACHES IN L ITERATURE

robotic arms [7], lower-limb exoskeletons [13], and elec- performance when applying CR up to 60%, which agrees with
tric wheelchair [12]. Kwak et al. [13] combined CCA with Tello et al. [25], who also proposed the use of CS for an
K -nearest neighbor (KNN) for SSVEP recognition, obtaining independent EEG SSVEP-based BCI, using CR of 75% on
high ACC like other approaches analyzed in Table V. This EEG epochs of 1 and 4 s to recognize two stimuli, obtaining
table also shows that our CS-ncCCA scheme improved the ACC of 72% and 85%, respectively. It is worth mentioning
TFS and latency values for 4271 kB and 273 ms, respectively, that our approach using a calibration stage (CS-wcCCA) also
outperforming other calibration-free systems. On the other performed well on 1 s EEG epochs, achieving ACC of 81% for
hand, our other CS-wcCCA approach achieved NTF and 40 stimuli. Although various research studies have developed
latency values of 2029 kB and 129 ms, respectively, which SSVEP–BCI systems [37], [38], [41], [42] to classify a high
is comparable to other results previously reported in studies number of targets (for instance in [42], with 160 targets,
that proposed systems with the stage of calibration. resulting in ACC of 87% and ITR of 78 b/min), we focused on
Chen et al. [7] achieved ACC of 97% and ITR of 17 b/min obtaining a solution to enhance greatly the ITR, maintaining
for four targets (left, right, forward, reverse). We believe that good accuracy in order to build a cloud-BCI for a robotic
using only these four commands in the next stage of our wheelchair.
research, both CS-ncCCA and CS-wcCCA schemes may also
increase the ACC, achieving further high ITR. It is worth V. C ONCLUSION
mentioning that our approaches reached ACC around 80% An SSVEP-based BCI using CS was developed here to
for classifying 40 SSVEP stimuli. Thus, these findings can convey control commands for a robotic wheelchair. We com-
be also translated to a robotic wheelchair, as done by Mistry bined data compression with the CCA method, exploring two
et al. [12], were similarly reported ACC of 80%. Our findings schemes for SSVEP recognition: calibration-free and with a
also agree with Chen et al. [41], which reported ACC of calibration stage. As a result, the best performance for both
80.1% by applying the calibration-free standard CCA over two-approaches was obtained when applying a compression
window lengths of 2 s during offline EEG processing and rate of 60% (ACC of 85% and ITR of 102 b/min for windows
SSVEP recognition, improving slightly their results in online of 2 s), and a compression rate of 62% (ACC of 85% and ITR
experiments (mean ACC of 84.1% and ITR of 105 b/min). of 195 b/min for windows of 1 s), respectively. The approaches
FBCCA has demonstrated great potential to increase the ITR with data compression produced good accuracy, increased
of SSVEP BCIs, as shown in Table V, which can be considered considerably the ITR, and reduced TFS, TCP and latency.
in future works, adding filter-banks in both CS-ncCCA and As a novelty, we explored CS without data reconstruction for
CS-wcCCA approaches. Notice that high ITR allows designing SSVEP-BCIs based on the CCA method using calibration-free
SSVEP-BCIs with more commands, as a recent study that used or a calibration stage. As a highlight, we found that different
FBCCA in a calibration-free BCI to decode 160 targets [42]. compression rates up to 60% on short EEG segments can
Likewise to previous research [7], [15], [25], [37], we found preserve the EEG information, guaranteeing high accuracy
that the BCI accuracy increased when increasing the win- and enhancing the ITR. The proposed methods reduce the
dow length for EEG processing and SSVEP recognition. For processing time, and the size of the file to be transferred and,
instance, the proposed CS-ncCCA and CS-wcCCA approaches consequently, reduce the latency in the communication. The
on 4 s window lengths reached ACC of 93% and 95%, communication system developed in ROS facilitates training
respectively, which decreased to 80% and 84% for shortest on a wheelchair simulator or a real wheelchair. The findings
EEG epochs of 2 s. Both approaches presented a consistent without a doubt may impact greatly a future centralized

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 RIVERA-FLOR et al.: CCA-BASED CS FOR SSVEP-BASED BCIs TO COMMAND A ROBOTIC WHEELCHAIR 4010510

cloud-BCI for commanding robotic wheelchairs. In particular, [17] K. A. Hernandez-Ossa et al., “Simulation system of electric-powered
relatively low-cost EEG wireless sensors can be used to wheelchairs for training purposes,” Sensors, vol. 20, no. 12, p. 3565,
2020.
transmit short-time EEG compressed data via the internet [18] P. Arpaia, L. Duraccio, N. Moccaldi, and S. Rossi, “Wearable brain–
network to a cloud-BCI, avoiding additional delays due to computer interface instrumentation for robot-based rehabilitation by
redundant information. A centralized cloud-BCI system in a augmented reality,” IEEE Trans. Instrum. Meas., vol. 69, no. 9,
pp. 6362–6371, Sep. 2020.
server also helps to reduce costs on the user terminal, as pow- [19] D. Caetano et al., “The augmented reality telerehabilitation system for
erful computers and software installation for heavy signal powered wheelchair user’s training,” J. Commun. Inf. Syst., vol. 35,
processing are not necessary anymore. Nextly both CS-ncCCA pp. 51–60, Mar. 2020.
[20] B. A. Jnr, “Use of telemedicine and virtual care for remote treatment in
and CS-wcCCA approaches will be translated in a cloud-BCI response to COVID-19 pandemic,” J. Med. Syst., vol. 44, no. 7, pp. 1–9,
to make possible the remote control of a wheelchair at distance Jul. 2020.
by novice users with severe motor impairments. Other issues, [21] M. V. R. Blondet, A. Badarinath, C. Khanna, and Z. Jin, “A wearable
real-time BCI system based on mobile cloud computing,” in Proc. 6th
such as delays produced by the wheelchair actuators, video Int. IEEE/EMBS Conf. Neural Eng. (NER). IEEE, 2013, pp. 739–742,
feedback, and real-time control will be addressed in future doi: 10.1109/NER.2013.6696040.
works to provide more reliable navigation. [22] A. Wang, F. Lin, Z. Jin, and W. Xu, “A configurable energy-efficient
compressed sensing architecture with its application on body sensor
networks,” IEEE Trans. Ind. Informat., vol. 12, no. 1, pp. 15–27,
R EFERENCES Feb. 2016.
[23] X. Liu, B. Subei, M. Zhang, A. G. Richardson, T. H. Lucas, and
[1] T. Bastos-Filho, A. Floriano, E. Couto, and R. J. Godinez-Tello, J. Van der Spiegel, “The PennBMBI: A general purpose wireless brain-
“Towards a system to command a robotic wheelchair based on indepen- machine-brain interface system for unrestrained animals,” in Proc. IEEE
dent SSVEP–BCI,” in Smart Wheelchairs and Brain-Computer Inter- Int. Symp. Circuits Syst. (ISCAS), Jun. 2014, pp. 650–653.
faces. Amsterdam, The Netherlands: Elsevier, 2018, pp. 369–379. [24] M. Balouchestani and S. Krishnan, “Robust compressive sensing algo-
[2] T. F. Bastos-Filho et al., “Towards a new modality-independent interface rithm for wireless surface electromyography applications,” Biomed.
for a robotic wheelchair,” IEEE Trans. Neural Syst. Rehabil. Eng., Signal Process. Control, vol. 20, pp. 100–106, Jul. 2015.
vol. 22, no. 3, pp. 567–584, Jun. 2014.
[25] R. G. Tello, J. K. Pant, S. M. Müller, S. Krishnan, and T. F. Bastos-
[3] R. Na et al., “An embedded lightweight SSVEP-BCI electric wheelchair Filho, “An evaluation of performance for an independent SSVEP-BCI
with hybrid stimulator,” Digit. Signal Process., vol. 116, Sep. 2021, based on compressive sensing system,” in World Congress on Medical
Art. no. 103101. Physics and Biomedical Engineering, June 7-12, 2015, Toronto, Canada
[4] D. Zhu, J. Bieger, G. G. Molina, and R. M. Aarts, “A survey of stimu- (IFMBE Proceedings), vol. 51. Springer Verlag, 2015, pp. 982–985, doi:
lation methods used in SSVEP-based BCIs,” Comput. Intell. Neurosci., 10.1007/978-3-319-19387-8_239.
vol. 2010, pp. 1–12, Jan. 2010.
[26] S. Sharma, S. Chaudhury, and D. Jayadeva, “Block sparse variational
[5] L. F. Nicolas-Alonso and J. Gomez-Gil, “Brain computer interfaces, a
Bayes regression using matrix variate distributions with application to
review,” Sensors, vol. 12, no. 2, pp. 1211–1279, 2012.
SSVEP detection,” IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 1,
[6] A. Floriano, D. Delisle-Rodriguez, P. F. Diez, and T. F. Bastos-Filho,
pp. 1–15, Jan. 2020.
“Assessment of high-frequency steady-state visual evoked potentials
[27] M. Rani, S. Dhok, and R. Deshmukh, “A systematic review of compres-
from below-the-hairline areas for a brain-computer interface based
sive sensing: Concepts, implementations and applications,” IEEE Access,
on depth-of-field,” Comput. Methods Programs Biomed., vol. 184,
vol. 6, pp. 4875–4894, 2018.
Feb. 2020, Art. no. 105271.
[7] X. Chen, B. Zhao, Y. Wang, and X. Gao, “Combination of high- [28] D. Gurve, D. Delisle-Rodriguez, T. Bastos-Filho, and S. Krishnan,
frequency SSVEP-based BCI and computer vision for controlling a “Trends in compressive sensing for EEG signal processing applications,”
robotic arm,” J. Neural Eng., vol. 16, no. 2, Apr. 2019, Art. no. 026012. Sensors, vol. 20, pp. 1–21, Jul. 2020.
[8] Z. Li, W. He, C. Yang, S. Qiu, L. Zhang, and C. Y. Su, “Teleoperation [29] M. Nakanishi, Y. Wang, X. Chen, Y. Wang, X. Gao, and T.-P. Jung,
control of an exoskeleton robot using brain machine interface and visual “Enhancing detection of SSVEPs for a high-speed brain speller using
compressive sensing,” in Proc. World Congr. Intell. Control Automat., task-related component analysis,” IEEE Trans. Biomed. Eng., vol. 65,
Sep. 2016, pp. 1550–1555. no. 1, pp. 104–112, Jan. 2018.
[9] L. Angrisani, P. Arpaia, N. Moccaldi, and A. Esposito, “Wearable [30] L. Liang et al., “Optimizing a dual-frequency and phase modulation
augmented reality and brain computer interface to improve human-robot method for SSVEP-based BCIs,” J. Neural Eng., vol. 17, Aug. 2020,
interactions in smart industry: A feasibility study for SSVEP signals,” Art. no. 046026.
in Proc. IEEE 4th Int. Forum Res. Technol. Soc. Ind. (RTSI), Sep. 2018, [31] C. M. Wong et al., “Learning across multi-stimulus enhances target
pp. 1–5. recognition methods in SSVEP-based BCIs,” J. Neural Eng., vol. 17,
[10] M. Xiong et al., “A low-cost, semi-autonomous wheelchair controlled Jun. 2020, Art. no. 016026.
by motor imagery and jaw muscle activation,” in Proc. IEEE Int. Conf. [32] C. M. Wong et al., “Inter- and intra-subject transfer reduces calibration
Syst., Man Cybern. (SMC), Oct. 2019, pp. 2180–2185. effort for high-speed SSVEP-based BCIs,” IEEE Trans. Neural Syst.
[11] B. He, S. Gao, H. Yuan, and J. R. Wolpaw, “Brain-computer inter- Rehabil. Eng., vol. 28, no. 10, pp. 2123–2135, Aug. 2020.
faces,” in Neural Engineering. USA: Springer, 2013, pp. 87–151, doi: [33] B. Liu, X. Chen, N. Shi, Y. Wang, S. Gao, and X. Gao, “Improving the
10.1007/9781461452270. performance of individually calibrated SSVEP-BCI by task-discriminant
[12] K. S. Mistry, P. Pelayo, D. G. Anil, and K. George, “An SSVEP based component analysis,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 29,
brain computer interface system to control electric wheelchairs,” in Proc. pp. 1998–2007, 2021.
IEEE Int. Instrum. Meas. Technol. Conf. (IMTC), May 2018, pp. 1–6. [34] Y. Ke, P. Liu, X. An, X. Song, and D. Ming, “An online SSVEP-
[13] N.-S. Kwak, K.-R. Müller, and S.-W. Lee, “A lower limb exoskeleton BCI system in an optical see-through augmented reality environment,”
control system based on steady state visual evoked potentials,” J. Neural J. Neural Eng., vol. 17, no. 1, Feb. 2020, Art. no. 016066.
Eng., vol. 12, Aug. 2015, Art. no. 056009. [35] L. Joseph and A. Johny, Programming With ROS. Berkeley, CA, USA:
[14] Y. Wang, X. Gao, B. Hong, C. Jia, and S. Gao, “Brain-computer Apress, 2022, pp. 173–240.
interfaces based on visual evoked potentials,” IEEE Eng. Med. Biol. [36] Z. Lin, C. Zhang, W. Wu, and X. Gao, “Frequency recognition based
Mag., vol. 27, no. 5, pp. 64–71, Sep. 2008. on canonical correlation analysis for SSVEP-based BCIS,” IEEE Trans.
[15] G. Ming, W. Pei, H. Chen, X. Gao, and Y. Wang, “Optimizing spa- Biomed. Eng., vol. 54, no. 6, pp. 1172–1176, Jun. 2007.
tial properties of a new checkerboard-like visual stimulus for user- [37] M. Nakanishi, Y. Wang, Y. Wang, Y. Mitsukura, and T. Jung, “A high-
friendly SSVEP-based BCIs,” J. Neural Eng., vol. 18, no. 5, Oct. 2021, speed brain speller using steady-state visual evoked potentials,” Int.
Art. no. 056046. J. Neural Syst., vol. 24, no. 6, Sep. 2014, Art. no. 1450019.
[16] A. Floriano, V. L. Carmona, P. F. Diez, and T. F. Bastos-Filho, “A study [38] X. Chen, Y. Wang, S. Gao, T.-P. Jung, and X. Gao, “Filter bank canon-
of SSVEP from below-the-hairline areas in low-, medium-, and high- ical correlation analysis for implementing a high-speed SSVEP-based
frequency ranges,” Res. Biomed. Eng., vol. 35, no. 1, pp. 71–76, brain–computer interface,” J. Neural Eng., vol. 12, no. 4, Aug. 2015,
Mar. 2019. Art. no. 046008.

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti
 4010510 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

[39] P. Yuan, X. Gao, B. Allison, Y. Wang, G. Bin, and S. Gao, “A study of Denis Delisle-Rodriguez received the B.E. degree
the existing problems of estimating the information transfer rate in online in telecommunication and electronic engineering and
brain–computer interfaces,” J. Neural Eng., vol. 10, no. 2, Apr. 2013, the M.Sc. degree in biomedical engineering from the
Art. no. 026014. University of Oriente, Santiago de Cuba, Cuba, in
[40] N. Japkowicz and M. Shah, Evaluating Learning Algorithms: A Classi- 2005 and 2013, respectively, and the Ph.D. degree
fication Perspective. Cambridge, U.K.: Cambridge Univ. Press, 2011. in electrical engineering from the Federal University
[41] X. Chen, Z. Chen, S. Gao, and X. Gao, “A high-ITR SSVEP-based BCI of Espírito Santo (UFES), Vitória, Brazil, in 2017.
speller,” Brain-Comput. Interfaces, vol. 1, nos. 3–4, pp. 181–191, 2014. He is currently a Professor at the Santos Dumont
[42] Y. Chen, C. Yang, X. Ye, X. Chen, Y. Wang, and X. Gao, “Implementing Institute, Macaíba, Brazil. His research interests
a calibration-free SSVEP-based BCI system with 160 targets,” J. Neural include biomedical signal processing, pattern recog-
Eng., vol. 18, Jun. 2021, Art. no. 046094. nition, brain–computer interfaces, and robotics for
[43] X. Chen, Y. Wang, M. Nakanishi, X. Gao, T.-P. Jung, and S. Gao, “High- rehabilitation and assistance.
speed spelling with a noninvasive brain–computer interface,” Proc. Nat.
Acad. Sci. USA, vol. 112, no. 44, pp. E6058–E6067, Nov. 2015.

Hamilton Rivera-Flor received the B.E. degree in

electronic engineering from Universidad del Valle,
Cali, Colombia, in 2008, and the M.Sc. degree in
electrical engineering from the Federal University
of Espírito Santo, Vitória, Brazil, in 2017, where he Ricardo Mello received the B.Sc., M.Sc., and the
is currently pursuing the Ph.D. degree. Ph.D. degrees in electrical engineering from the Fed-
In 2018, he was a Visiting Scholar at Ryerson Uni- eral University of Espírito Santo (UFES), Vitória,
versity, Toronto, ON, Canada. His research interests Brazil, in 2015, 2018, and 2020, respectively.
include biomedical signal processing, mobile robots, In 2020, he joined the Department of Elec-
cloud robotics, the Internet of Things (IoT), and trical Engineering, UFES, as an Assistant Pro-
assistive technology.
fessor. He developed part of his thesis at the
Colombian School of Engineering Julio Garavito
(ECIJG), Bogotá, Colombia, and at the Centre for
Dharmendra Gurve (Member, IEEE) received Automation and Robotics (UPM-CSIC), Madrid,
the B.E. degree from Rajiv Gandhi Proudyogiki Spain. His research interests include cloud robotics,
Vishwavidyalaya, Bhopal, India, in 2012, the autonomous systems, assistive robotics, and human–robot interaction.
M.Tech. degree from the Dr. B. R. Ambedkar
National Institute of Technology Jalandhar,
Jalandhar, India, in 2015, and the Ph.D. degree
from Ryerson University, Toronto, ON, Canada,
in 2020.
In 2020, he joined McGill University, Montreal,
QC, Canada, as a Computational Neuroscientist
and worked on the cognitive control of auditory
environment project that combines neuroscience, advanced acoustics, and
Artificial Intelligence (AI) to design an assistive device that is steered by Teodiano Bastos-Filho received the B.E. degree
brain activity. He is currently working at Ryerson University as a Project in electrical engineering from Universidade Federal
Scientist, where he is developing an on-device machine learning (ML) and do Espírito Santo, Vitória, Brazil, in 1987 and the
deep learning (DL) framework. His current research interests include Tiny Ph.D. degree in physical sciences (electricity and
ML for healthcare applications. electronics) from the Universidad Complutense de
Madrid, Madrid, Spain, in 1994.
He was a Post-Doctoral Fellow at the University
Alan Floriano received the B.Sc. degree in com- of Alcalá, Alcalá de Henares, Spain, in 2005 and
puter engineering and the M.Sc. and the Ph.D. at RMIT University, Melbourne, VIC, Australia,
degrees in electrical engineering from the Federal in 2012. He is a Consultant of the Fund for Scientific
University of Espírito Santo (UFES), Vitória, Brazil, Research (FNRS) and Editorial Board Member of
in 2014, 2016, and 2019, respectively. the journal Scientific Reports – Nature. He has more than 600 scientific
He is a Post-Doctoral Fellow at the Institute of publications (H-index = 30 at Scholar Google, H-index = 24 at Scopus, and
Mathematics and Computer Sciences, University of H-index = 18 at Web of Science) and experience in electrical and biomedical
São Paulo, São Paulo, Brazil. His research interests engineering, focusing on electronic systems of measure and control, acting on
include signal processing, machine learning (ML), the following subjects: control, sensors, mobile robots, rehabilitation robotics,
and human–computer interaction. biomedical signal processing (surface electromyography (sEMG), electroocu-
lography (EOG), electrocardiogram (ECG), and EEG), and bioinformatics.

licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on November 28,2024 at 11:57:58 UTC from IEEE Xplore. Restricti