Received 8 March 2023, accepted 6 April 2023, date of publication 11 April 2023, date of current version 14 April 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3266306

Employment of Domain Adaptation Techniques

in SSVEP-Based Brain–Computer Interfaces
ANDREA APICELLA 1 , PASQUALE ARPAIA 2 , (Senior Member, IEEE),
EGIDIO DE BENEDETTO 1 , (Senior Member, IEEE),
NICOLA DONATO 3 , (Senior Member, IEEE), LUIGI DURACCIO4 ,
SALVATORE GIUGLIANO 1 , AND ROBERTO PREVETE 1
1 Department of Electrical Engineering and Information Technology, University of Naples Federico II, 80125 Naples, Italy
2 InterdepartmentalResearch Center in Health Management and Innovation in Healthcare, University of Naples Federico II, 80125 Naples, Italy
3 Department of Engineering, University of Messina, 98122 Messina, Italy
4 Department of Electronics and Telecommunications, Polytechnic University of Turin, 10129 Turin, Italy

Corresponding author: Egidio De Benedetto ([email protected])

This work was supported in part by the European Union-FSE-REACT-EU, PON Research and Innovation 2014-2020 DM1062/2021 under
Contract 18-I-15350-2; and in part by the Ministry of University and Research, PRIN research project ‘‘BRIO–BIAS, RISK, OPACITY in
AI: design, verification and development of Trustworthy AI.’’, under Project 2020SSKZ7R.

ABSTRACT This work addresses the employment of Machine Learning (ML) and Domain Adaptation (DA)
in the framework of Brain-Computer Interfaces (BCIs) based on Steady-State Visually Evoked Potentials
(SSVEPs). Currently, all the state-of-the-art classification strategies do not consider the high non-stationarity
typical of brain signals. This can lead to poor performance, expecially when short-time signals have to be
considered to allow real-time human-environment interaction. In this regard, ML and DA techniques can
represent a suitable strategy to enhance the performance of SSVEPs classification pipelines. In particular,
the employment of a two-step DA technique is proposed: first, the standardization of the data per subject
is performed by exploiting a part of unlabeled test data during the training stage; second, a similarity
measure between subjects is considered in the selection of the validation sets. The proposal was applied
to three classifiers to verify the statistical significance of the improvements over the standard approaches.
These classifiers were validated and comparatively tested on a well-known public benchmark dataset.
An appropriate validation method was used in order to simulate real-world usage. The experimental results
show that the proposed approach significantly improves the classification accuracy of SSVEPs. In fact,
up to 62.27 % accuracy was achieved also in the case of short-time signals (i.e., 1.0 s). This represents a
further confirmation of the suitability of advanced ML to improve the performance of BCIs for daily-life
applications.

INDEX TERMS Brain–computer interface, domain adaptation, EEG, EEGNet, health 4.0, instrumentation,
machine learning, neural engineering, neural networks, SSVEP, real-time systems.

I. INTRODUCTION non-invasive, cheap, and is endowed with high temporal res-

In recent years, human-machine interaction has been olution to allow real-time operation [3], [4]. In fact, different
significantly improved by the widespread diffusion of EEG-based BCI paradigms, such as P300 [5] and Motor
Brain-Computer Interfaces (BCIs) [1]. BCIs are an emerging Imagery [6], [7], [8], [9], [10] have already been success-
technology integrating hardware and software to create a fully employed in several contexts but, in particular, Steady-
direct communication pathway between the human brain and State Visually Evoked Potentials (SSVEPs) have gained
external devices [2]. Among the different ways of decoding outstanding relevance for the development of applications
brain activity, Electroencephalography (EEG) is receiving in healthcare, [11], [12], [13] entertainment [14], and indus-
a strong interest by the scientific community since it is try [15], [16] owing to quick response, easy detection, high
signal-to-noise ratio (SNR) [17]. As a matter of fact, the
The associate editor coordinating the review of this manuscript and classification of SSVEPs can be performed with good results
approving it for publication was Gang Wang . even with simple, trainingless algorithms, such as Power
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 36147
 A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

Spectral Density Analysis (PSDA) or Canonical Correla- training data. Moreover, also in this case the Dataset Shift
tion Analysis (CCA) [18]. Nevertheless, there is much room problem can arise insofar as data are collected from substan-
for improvements aimed at employing SSVEP-based BCIs tially different sessions.
in challenging contexts, where the requirements are very For these reasons, newer studies tried to overcome the
demanding [19], [20]. Dataset Shift problem in EEG-based BCIs [33]. In particular,
To this aim, the widespread adoption of 4.0 enabling tech- Domain Adaptation (DA) strategies try to construct models
nologies such as Artificial Intelligence (AI) and, in particular, able to generalize on unseen data exploiting knowledge given
Machine Learning (ML) [21], has raised the question of by available unlabelled data. These strategies rely on the
whether this family of technologies can improve the per- assumption that a small initial part (even unlabeled) of the
formance of such systems [22]. Among the ML paradigms, new user’s data is already available before the actual classifi-
supervised ML models learn to predict outputs on the basis cation. Although time-consuming, acquiring unlabeled data
of given examples of relationships between input data and from a new user/session is still easier than getting labeled
outputs (the training data). In this regard, several related data, leading a BCI system to be more comfortable than one
works have already addressed the use of ML classifiers in trained only on user/session-specific labeled data.
SSVEPs classification, such as (i) Support Vector Machine
(SVM) [23], (ii) k-Nearest Neighbors (k-NN) [24], and
(iii) Artificial Neural Networks (ANNs) [25], by showing B. PROPOSAL
that ML represents a very promising strategy to boost the Starting from these considerations, in this work, a DA tech-
performance of their SSVEP-based BCIs. It has also been nique is proposed in the framework of the SSVEPs classifi-
empirically shown that increasing the model complexity can cation. More specifically, a two-step DA method is validated
lead to further enhancements in classification accuracy in on a public dataset described in [34] and composed of 35 sub-
several cases. For the sake of example, in [26] and [27], the jects and 40 simultaneous flickering stimuli. The remarkable
realized Convolutional Neural Networks (CNNs) allowed to number of subjects and flickering stimuli allows considering
achieve, under specific conditions, higher performance than this dataset a challenging benchmark to significantly test
Filter Bank Canonical Correlation Analysis (FBCCA) [28], SSVEPs classification algorithms. Following [27], a suitable
which is currently considered the state of the art about validation method was adopted to simulate real-world usage
SSVEPs classification [27]. in a statistically significant way.
However, when short-time input signals are taken into The proposed DA technique is composed of two main
account (i.e., lower than 1.5 s), the classification of SSVEPs steps, the former consisting in a per-subject z-score nor-
can still be strongly improved as a means to allow real- malization (instead of the classical z-score applied on all
time human-environment interaction. Currently, in fact, most the data without regard to the belonging domain), and the
of the literature considers SSVEPs only as a steady-state latter consisting in a simple change in the classical neural
response given by purely oscillatory components synchro- network training procedure. More in detail, in the classi-
nized in phase with the stimulation source, thus not con- cal neural network training stage a subset of the training
sidering the intrinsic non-stationarity of EEG signals [29], data (named validation set) is used to prevent the network
resulting in strong differences across the EEGs acquired from to overfit on the training data. Then, the performance of
different subjects or at different time intervals (sessions). This the network on the validation data are computed at the end
can represent a limitation in SSVEPs classification and, in our of each training iteration, and the training stage is stopped
knowledge, it is currently taken into account only in few when the validation performance starts to degrade rather than
studies [30], [31]. improve [35]. However, the validation set is usually selected
randomly from the training data, without any regards about
A. THE DATASET SHIFT PROBLEM the original distribution of the data. In cases where several
From a ML point of view, this issue can be viewed as an distribution are involved, such as in EEG data acquired from
instance of the Dataset Shift problem [32]. In a nutshell, several subjects, this could not be the better choice, since
a Dataset Shift arises when the distribution of the data used data too dissimilar from the test data could be chosen, not
to train the ML model differs from the data distribution leading the network to generalize toward unseen domains.
used outside of the training stage. This violates one of the Instead, if also a small part of unlabelled data coming from the
main assumption of ML approaches, stating that all the data, target subject are available during the training time, this can
no matter if involved in training or not, come from the be used to choose the validation set in a smarter way respect
same probability distribution. As a consequence, the trained to the random choice. In this study, 20 % of the test data
model won’t work as expected on data acquired from dif- was considered as calibration data, therefore available at the
ferent subjects or at different sessions with respect to the training data without any label. These data can be used both
ones used during the training stage. This problem is usually to standardize the data and to compute a similarity measure
mitigated by training specific models for each subject (called between subjects to select the validation sets.
intra − subject models). However, an intra-subject model can The statistical significance of the improvements made by
be used only on data acquired from the subject providing the proposed DA technique over the standard approaches

36148 VOLUME 11, 2023

A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

was validated on three different neural network clas-

sifiers: ShallowConvNet, DeepConvNet, and EEGNet.
ShallowConvNet and DeepConvNet are two CNNs devel-
oped in [36] and used in Motory Imagery-based BCIs.
Instead, EEGNet is a compact CNN successfully employed
in several tasks involving different types of EEG sig-
nals, such as P300 visual-evoked potentials, Error-Related
Negativity Response (ERN), Movement-Related Cortical
Potentials (MRCP), and Sensory Motor Rhythms (SMR),
showing comparably high performance with respect to the
reference algorithms [37]. More recently, it has gained inter-
est also in SSVEP-based BCIs [38]. In particular, promis-
ing results in terms of classification accuracy are reported
in literature [39]. In our knowledge, however, this is the
first time that EEGNet is tested to the benchmark dataset
considered.
The paper is organized as follows. Sec. II provides a
background on the SSVEPs classification problem, reporting FIGURE 1. A 10-Hz SSVEP in time domain (a) and frequency domain (b).

some of the most widely adopted processing strategies over

the years. Then, Sec. III describes, in brief, the classifiers
chosen, along with the validation method and the DA tech-
nique considered. Therefore, Sec. IV reports the experimental
setup, and Sec. V provides a comparative analysis between
the results obtained by the proposed approach and those
achieved with traditional, state-of-the-art strategies. Finally,
conclusions are drawn.

II. BACKGROUND AND RELATED WORKS

This Section provides a background on the SSVEPs classifi-
cation problem. In addition, an overview on some of the most
widely adopted processing strategies is given.

A. THE SSVEPs CLASSIFICATION PROBLEM

SSVEPs are exogenous brain potentials [40], elicited in
the primary visual cortex when a flickering stimulus is FIGURE 2. Typical architecture of a SSVEP-based BCI.
observed by the user. Stimulation frequency bands of the
visual stimuli usually range from 6 Hz to 30 Hz, although
the best Signal to Noise Ratio (SNR) is achieved in the captures the user’s brain signals, which are digitized and
range 8 ÷ 15 Hz [41]. The physiological SSVEP brain processed by means of a SSVEP classification algorithm.
response is typically inducted after a latency ranging The aim of this algorithm is to deduce which stimulus
from 80 to 160 ms [42]. It is a sinusoidal-like waveform, has been observed by the user. Therefore, the recognition
composed of a fundamental frequency equal to that of the of N stimuli at different frequencies can be viewed as
gazed stimulus, and often higher harmonics [43], as shown a N -class classification problem. Once the classifica-
in Fig. 1. tion has been made, an output command is sent to the
In practical applications, stimuli at different frequencies BCI Application, which provides a feedback to the user
are simultaneously displayed to the user. Each stimulus is regarding the selection performed. Clearly, if the classifica-
associated to a specific command: the user, by looking at tion is successful, the output command corresponds to the
the desired flickering stimulus, is able to send the related choice desired by the user.
command to the target application.
A representative architecture of SSVEP-based BCIs is B. PROCESSING STRATEGIES FOR CLASSIFYING SSVEPs
shown in Fig. 2. A Stimuli Source (typically, a LCD monitor, An overview of some of the typical processing strategies for
or an eXtended Reality headset [16], [20]) is used to display N classifying SSVEPs is given as follows. In particular, a dis-
concurrent flickering stimuli. Each stimulus flickers at a dif- tinction between non-ML, hybrid, and ML methods is made.
ferent frequency from the others and is associated to a specific For the sake of simplicity, but without loss of generality,
command to send to the BCIapplication. An EEG headset a single-channel EEG signal is considered.

VOLUME 11, 2023 36149

 A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

1) NON-ML METHODS with given threshold values: the signal was marked
• PSDA: Since SSVEPs are characterized by frequency as classified only if the chosen correlation coefficient
peaks consistent with the observed flickering stimuli, the exceeded the thresholds. The classification performance
most intuitive approach used to detect and classify the achieved with the use of CCA are typically better than
elicited SSVEPs is based on a Power Spectral Density PSDA [44]. However, a band pass filtering for the EEG
Analysis (PSDA) [41]. This method is composed by can be often necessary during the pre-processing phase,
three steps: first, a Fast Fourier Transform (FFT) is due to the effect of spontaneous EEG activities not
applied to the user EEG; then, a PSD is performed in involved in SSVEP events.
the neighborhood of each of the N frequencies rendered • FBCCA: The FBCCA method is an enhancement of
on the display, and eventually its multiple m harmonics, CCA [28] and consists of three major procedures: (i) fil-
according to (1). ter bank analysis; (ii) CCA between SSVEP sub-band
  components and sinusoidal reference signals; and
m jkX n +k
1 X
2 (iii) signal classification. First, sub-band decomposi-
P(fn ) =  c(j) A (i) (1) tions are performed by the filter bank analysis by means
2mk + 1
j=1 i=jkn −k
of multiple filters with different pass-bands. In this way,
where: P(fn ) is the PSD coefficient for the given fre- the sub-band components XSBj (j = 1, 2, . . . , s) from
quency fn (n = 1, 2, . . . , N ), kn is the corresponding the original EEG X are obtained. After the filter bank
bin in frequency domain, k is the number of nearest analysis, the standard CCA is applied to each of the
bins to be considered, m is the number of chosen har- sub-band components separately. This results in cor-
monics, A is the signal amplitude, and c is a weight relation values between the sub-band components and
assigned to each harmonics. Finally, the classification the sinusoidal reference signals corresponding to the
is usually performed based on the hypothesis that the stimulation frequencies (n = 1, 2, . . . , N ). A correlation
observed stimulus is very likely to be the one with the value ρjn is obtained for each frequency n and each sub-
highest PSD [44]. The main drawback of PSDA is that band j according to (3).
it requires a minimum time window Tmin of the acquired h i
ρ n = ρ1n , ρ2n , . . . , ρjn , . . . , ρsn (3)
EEG in order to correctly discriminate the harmonics,
since an appropriate frequency resolution 1f = Tmin 1
is A weighted sum of squares of the correlation values
needed [45]. corresponding to all sub-band components is calculated
• CCA: An alternative way to process SSVEPs is as the feature for signal classification.
the Canonical Correlation Analysis (CCA) in time s
X
domain [16]. It is a multivariate statistical method of ρ̃ n = w(j) · (ρjn )2 (4)
correlating linear relationships between two sets of j=1
data [46]. CCA is performed between the EEG data D where j is the index of the sub-band. As the SNR of
and a set of sine waves yn (t) having the same frequencies SSVEP harmonics decreases as the response frequency
of the N stimuli rendered on the display, and eventually increases, the weights for the sub-band components are
their multiple harmonics. Given a frequency fn and the defined as follows:
number of harmonics m to consider, the set of sinewaves
yn (t) (n = 1, 2, . . . , N ) can be obtained according to (2). w(j) = j−a + b (5)
 
sin(2π fn t) where a and b are constants that maximize the clas-
 cos(2π fn t)
  sification performance. Therefore, N features ρ̃ n are

 sin(4π fn t)
 obtained (one for each frequency). Finally, the signal
  classification is performed on the basis that the observed
yn (t) =  cos(4π fn t) (2) frequency fz (z ∈ 1, . . . , N ) is that corresponding to
 

. . . the feature ρ̃ z with the maximum value. Typically, the

 

 
 sin(mπ fn t)
  number of filters s, the number of harmonics m, and the
values of a and b are determined using a grid search
cos(mπ fn t)
method in an offline analysis. A widely adopted practice
For each stimulus frequency fn , a correlation coefficient is to vary s from 1 to 10, m from 1 to 6, a from 0 to 2,
ρn is extracted by means of the CCA between D and and b from 0 to 1.
yn (t). Therefore, these coefficients are used for SSVEP
classification. For the sake of example, in [46] the output 2) HYBRID METHODS
of the classification was associated to the frequency At the state of the art, CCA-based algorithms provide the best
with the highest correlation coefficient extracted. Alter- performance in terms of classification accuracy [34], [49].
natively, in [1], [47], and [48] the maximum value However, recent works [50] showed that, for low-channels
among the correlation coefficients ρn was compared and low-stimuli setups, the adoption of hybrid approaches,

36150 VOLUME 11, 2023

A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

FIGURE 3. Main blocks of the three classifiers chosen: ShallowConvNet (left), DeepConvNet (center), and EEGNet (right).

based on a pre-processing of the EEG signal in time and Batch Normalization layer with ELU activation func-
frequency domains, and a ML-based classification, allows to tion, an Average Pooling layer, a Dropout layer, and,
outperform the results obtained by CCA. In particular, the finally, a Dense Softmax classification layer.
algorithm developed in [50] (named Features Reduction) is 2) DeepConvNet is a Convolutional Neural Network
constituted by (i) a pre-processing step, based both on Power developed in [36] and composed of four blocks:
Spectral Density Analysis (PSDA) and Canonical Correlation • The first block consists of two convolutional layers
Analysis (CCA), to extract significant features from the digi- which perform a temporal convolution and a spa-
tized EEG signal, and (ii) a classification step which employs tial filter. Then, Batch Normalization is applied,
ML classifiers such as SVM, K-NN, and shallow NN to followed by Exponential Linear Unit (ELU) [53]
classify the extracted features into an output z (z ∈ 1, . . . , N ), activation function. Finally, a Max Pooling layer
where N is the number of stimulation frequencies. and Dropout layer are used.
• The other three blocks are classic convolutional
3) ML METHODS
blocks, in which only one convolutional layer is
PodNet is a CNN developed by Podmore et al. [27]. It is followed by the same layers of the first block:
constituted by several blocks (Pods), each one made up of a Batch Normalization, Activation, Max Pooling,
Convolutional layer, a Drop-out layer, a Batch Normalization and Dropout.
layer, a Rectifier Linear Unit (ReLU) layer, and a Max Pooling
3) EEGNet is a Convolutional Neural Network origi-
layer. The final Pod contains a dense layer which outputs
nally designed to be applied to a wide variety of BCI
to a Softmax operation to classify the EEG into one of the
paradigms (for further details, please refer to [37]).
possible z (z ∈ 1, . . . , N ) classes, where N is the number of
A brief description of the main blocks is provided as
stimulation frequencies.
follows:
All network weights are initialized using the Xavier
• In the first block, a classical convolutional layer
method [51] and updated following the Adam optimization
algorithm [52]. In [27], it has been shown that PodNet man- is applied to the input EEG signal, followed
ages to outperform FBCCA, even if only under specific by a Depthwise Convolution layer [37]. Then,
conditions (i.e., low-volume EEG electrode arrangements). Batch Normalization is applied followed by ELU.
Nevertheless, there is still much room for improvements Finally, an Average Pooling layer of size (1, 4)
when short-time input signals are taken into account, which is used to reduce the sampling rate of the signal
are critical for facilitating real-time human-environment followed by a Dropout layer.
• In the second block, a Separable Convolution [37]
interaction.
is used. After a Batch Normalization followed by
III. MATERIALS AND METHODS the ELU activation function, a final Average Pool-
This Section describes the classifiers chosen, along with the ing layer followed by a Dropout layer is applied.
validation method and DA technique proposed. • Final classification is made using applying a Soft-
max function directly, avoiding the use of a prior
A. CLASSIFIERS CHOSEN dense layer for feature aggregation to reduce the
The three classifiers employed in this work were Shallow- number of free parameters in the model.
ConvNet, DeepConvNet, and EEGNet, as shown in Fig. 3.
1) ShallowConvNet is a Convolutional Neural Network B. VALIDATION METHOD
developed in [36] and characterized by a single block, EEGNet was validated on the benchmark dataset with a
where two convolutional layers, which perform a tem- repeated Hold-Out validation: following [27], this work
poral convolution and a spatial filter, are followed by a included 25 subjects in the training set, 5 in the validation set,

VOLUME 11, 2023 36151

 A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

and the remaining 5 in the test set. In addition, 10 runs were A. DATA DESCRIPTION
made for each combination of the model’s hyperparameters to The benchmark dataset has the following features:
make the reported results more statistically significant. Only • Subjects: 35 healthy subjects (17 females and 18 males,
novel BCI subjects were included in the test set to simulate aged 17-34 years, mean age: 22 years), having nor-
real-world usage, as proposed in [27]. mal or corrected-to-normal vision, participated in this
study. 8 subjects had previous experience in SSVEP-
C. DOMAIN ADAPTATION TECHNIQUE based BCI.
A two-step Domain Adaptation technique was applied during • Stimulus Presentation: An offline BCI experiment using
the training of the three analyzed models (ShallowConvNet, a 40-target BCI speller was designed. The 5 × 8 stimulus
DeepConvNet, and EEGNet) on the benchmark dataset: matrix was presented on a 23.6-in LCD monitor (Acer
1) DA-based Standardization: The relying hypothesis is GD245 HQ, response time: 2 ms) with a resolution
that each subject data can be considered as belonging of 1920 × 1080 pixels, and a refresh rate of 60 Hz.
to a different Domain. Therefore, instead of the clas- The viewing distance to the screen was 70 cm. The
sical z-score normalization on the whole training data, sizes of stimulus and character were 140 × 140 and
a z-score standardization of the data subject-by-subject 32 × 32 pixels square, respectively. The size of the
was made. In particular, for each subject Si belonging whole matrix area was 1510 × 1037 pixels. Both the
to the training and validation sets, all the related data vertical and horizontal distances between two neighbor-
were used to extract the subject mean value µi and the ing stimuli were 50 pixels. The stimulus program was
subject standard deviation σi . Therefore, a subject-wise developed under MATLAB using the Psychophysics
z-score normalization was performed. Instead, for each Toolbox Ver. 3. The 40 characters were coded using a
subject St belonging to the test set, 20 % of the data joint frequency and phase modulation (JFPM) approach.
was considered as unlabelled calibration data available In particular, the chosen frequencies were in the range
during the training. The mean µt and the standard [8.0-15.8] Hz with a 0.2 Hz step, while the phase
deviation σt was computed on the calibration data and values had a 0.5 π step. A sampled sinusoidal stim-
used to perform a z-score normalization on the whole ulation method was applied to present visual flick-
subject data. ers on the LCD monitor. Given a frequency f and a
2) Similarity between Subjects: The similarity between phase θ, the stimulus sequence s(f , θ, i) was generated
subjects was considered in the selection of the valida- by modulating the luminance of the screen according
tion sets. In general, a validation set is used to control to 6 [34]:
the overfit of the network during the training. More 1
in detail, the performance on the validation set are s(f , θ, i) = {1 + sin[2π f (i/RR) + θ]} (6)
2
computed during the training stage and if it tends to
get worse, the training is stopped. Differently from where sin(·) generates a sine wave, and i indicates the
classical neural networks’ learning strategies, where frame index in the sequence. The refresh rate of the LCD
the validation set is selected from the training data ran- monitor is indicated with RR. In the obtained stimulus
domly, in this study the validation sets were composed sequence, 0 and 1 represent darkest and highest lumi-
of the subjects closest to those who provided calibration nance, respectively.
data. Our starting hypothesis is that a proper validation • Data acquisition: A Synamps2 EEG acquisition unit
set selected on the basis of the calibration data provided (Neuroscan, Inc.) was used to record EEG data at a
by test subjects can lead to learn a model able to better sampling rate of 1 kSa/s. 64 electrodes, according to
classify the data coming from the target distribution. international 10-20 system [55], were used to record
In this study, Kullback-Leibler (KL) divergence [54] whole-head EEG. The reference electrode was placed
was used to measure the similarity between training at the vertex of the user scalp (Cz). The electrode
and calibration data. More in detail, for each pair of impedances were kept below 10 k  during the record-
subjects, their similarity is computed as follows: ing phase. A notch filter at 50 Hz was applied to remove
the power-line noise. For each subject, the experiment
K
1 X included six blocks. Each block was composed of 40 tri-
sim(Si , Sj ) = KL(Sik , Sjk ), ∀1 ≤ i, j ≤ N als, corresponding to all 40 squares. Each trial started
K
k=1 with a 0.5-s target cue. Subjects were asked to shift
where: S1 , S2 , . . . , SN are the N subjects and K is the their gaze to the indicated target as soon as possible.
number of EEG channels. After the cue, all stimuli started to flicker on the screen
concurrently for 5 s. Then, the screen became blank
IV. EXPERIMENTAL SETUP for 0.5 s, before the start of the next trial. Subjects were
This section provides an overview on the benchmark dataset asked to avoid eye blinks during the 5-s stimulation dura-
considered [34]. Moreover, the pre-processing strategies and tion. A rest for several minutes between two consecutive
the model selection criteria are described. blocks was foreseen.

36152 VOLUME 11, 2023

A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

B. PRE-PROCESSING TABLE 2. Optimized hyperparameters and variation ranges for

DeepConvNet.
Ten channels from the occipital and parietal areas were
selected for the experiments, namely PO8, PO7, PO6, PO5,
PO4, PO3, POz, O2, O1, and Oz, according to the Interna-
tional 10-20 System [34], as conducted in [27]. The time
windows considered were 0.5, 1.0, and 1.5 s as they are
typically considered the most challenging in view of using
BCIs for real-time applications [56]. Therefore, the analy-
sis at 5.0 s was excluded, also because the user would too
long be obliged to gaze at the desired flickering stimulus,
resulting in ocular discomfort. In all cases, the time dura-
tion of the target cue was discarded. The data extracted
TABLE 3. Optimized hyperparameters and variation ranges for EEGNet.
are then filtered by means of a band-pass Finite Impul-
sive Response (FIR) filter [50] with linear phase response,
which avoids distortions on the original data and preserves
the information contribution of the alpha and beta bands.
Therefore, the standardization of the data was performed
in two ways: (i) canonical z-score normalization [57], and
ii) according to the proposed DA technique mentioned
in Sec. III-C.

C. SELECTION OF THE BEST MODEL

A random search strategy [58] was adopted to select the time windows T , namely T = {0.5 s, 1.0 s, 1.5 s}. To val-
hyperparameters values of ShallowConvNet, DeepConvNet, idate the proposed approach, also a standard ML approach
and EEGNet. The hyperparameters spaces of each model are consisting in standard z-score normalization parameters com-
reported respectively in Tables 1, 2 and 3. In particular, on the puted on the whole training data and classical validation set
three models, the Temporal Kernel Length hyperparameter is composed of random data extracted from the training data
related to the signal sample rate and the time window input; was considered as baseline. In other words, the following two
with regards to EEGNet, the 2D convolutional layers of the experiments were made:
first and second blocks, share the same number of filters.
• DA experiment: in this case, the two-step DA technique
Furthermore, during the learning phase, a stop criterion was
proposed in Sec. III (DA-based standardization and sim-
used by means of a patience of 10 epochs. The hyperpa-
ilarity between subjects) is employed.
rameters values leading the model toward the best perfor-
• Standard experiments: in this case, a standard ML
mance in terms of mean classification accuracy (defined
approach (canonical z-score normalization and random
as the percentage of signals correctly classified) were
validation sets) is adopted.
selected.
As specified in Sec. III, the validation method used was an
TABLE 1. Optimized hyperparameters and variation ranges for Hold-Out with a 25-5-5 split repeated 10 times. This setup
ShallowConvNet. allows to compare the proposal with the results achieved
by PodNet reported in [27]. Moreover, the same validation
method was also applied to CCA and FBCCA, implemented
as described in [28]. However, several details needed for
experimental repeatability, such as the precise validation
procedure, are missing in the original study, therefore the
FBCCA results obtained in this work are different from
those reported in [28]. As stated in Sec. III, the 10 test sets
were created randomly, with the only condition of excluding
BCI-experienced subjects to simulate real-world usage. Pre-
liminary experiments, in fact, showed that the performance
V. EXPERIMENTAL RESULTS of the expert participants were significantly higher than those
In this Section, the results achieved by the proposed approach obtained by naive ones. For the sake of example, given
on the benchmark dataset [34] are reported. A comparison a 1.0-s time window, the mean accuracy obtained by
with the performance obtained by several well-known strate- EEGNet - DA on the eight experienced subjects was greater
gies proposed in literature, that are PodNet [27], CCA, and than 75 %, while that on the novel ones was only about 61 %.
FBCCA [28], was made. More in detail, Tab. 4 provides the A similar condition happened with FBCCA: in that case,
classification accuracy and standard deviations for several the mean accuracy obtained on the expert subjects was

VOLUME 11, 2023 36153

 A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

FIGURE 4. Boxplot of the obtained results as a function of the time window.

TABLE 4. Classification accuracy as a function of the time window T . The and ShallowConvNet when the two-step DA technique is
three classifiers chosen are shown in bold.
applied results greater than the baseline in all the analyzed
cases. For the sake of example, EEGNet - DA achieves
about 62 % accuracy at 1.0 s, while EEGNet - Standard
is limited to about 57 %, and FBCCA reaches only 39 %.
Therefore, the adoption of the proposed DA method leads
to an improvement of the performance with respect to the
baseline.
Fig. 4 illustrates the boxplot of the accuracy achieved
by all the aforementioned processing strategies, as a func-
tion of the time window T . The length of the whiskers is
set to 1.5 times the Interquartile Range (IQR). The statis-
tical significance of the results was tested by means of the
Paired T − Test.
• First, it was verified that the proposed DA tech-
nique significantly outperforms traditional ML strate-
gies. To this aim, the chosen null hypothesis H0 was that
the classifiers DA and Standard belonged to the same
population.
• Second, it was verified that the classifiers employed
with the DA technique significantly outperform FBCCA,
which is the gold standard in terms of SSVEPs classi-
fication. Therefore, the chosen null hypothesis H0 was
that the classifiers DA and FBCCA belonged to the same
population.
In Tab. 5, the details of the tests performed are shown
in terms of P-Value, that is the probability of failing to
reject the null hypothesis H0 . The obtained P-Values can be
considered acceptable to confirm that the adoption of DA
about 54 %, while that on the naive ones was only 40 %. techniques leads to relevant improvements in classification
Therefore, the proposed experiments including only accuracy with respect both to traditional ML approaches
BCI-novels subjects were considered more suitable to vali- and to FBCCA strategy, which is currently the state of the
date the usage of BCIs in real-world scenarios. art in the field of SSVEPs classification. These improve-
With regard to the findings reported in Tab. 4, the mean ments are more evident when the time windows are equal
classification accuracy achieved by EEGNet, DeepConvNet, to 0.5 s and 1.0 s. In fact, in these cases, the obtained

36154 VOLUME 11, 2023

A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

TABLE 5. Details about the p-values obtained from the t-tests conducted.

P-Values considerably decrease with respect to those [3] D. J. McFarland and J. R. Wolpaw, ‘‘Brain-computer interfaces for com-
obtained at 1.5 s. munication and control,’’ Commun. ACM, vol. 54, no. 5, pp. 60–66, 2011.
[4] R. A. Ramadan and A. V. Vasilakos, ‘‘Brain computer interface: Control
signals review,’’ Neurocomputing, vol. 223, pp. 26–44, Feb. 2017.
VI. CONCLUSION [5] R. Fazel-Rezai, B. Z. Allison, C. Guger, E. W. Sellers, S. C. Kleih,
In this work, the employment of Machine Learning and A. Kübler, ‘‘P300 brain computer interface: Current challenges and
emerging trends,’’ Frontiers Neuroeng., vol. 5, p. 14, Jul. 2012.
(ML) and Domain Adaptation (DA) in the framework of
[6] J. Jiang, C. Wang, J. Wu, W. Qin, M. Xu, and E. Yin, ‘‘Temporal combi-
Brain-Computer Interfaces (BCIs) based on Steady-State nation pattern optimization based on feature selection method for motor
Visually Evoked Potentials (SSVEPs) was addressed. Three imagery BCIs,’’ Frontiers Hum. Neurosci., vol. 14, p. 231, Jun. 2020.
well-known classifiers were employed to validate the pro- [7] Y. Pei, Z. Luo, Y. Yan, H. Yan, J. Jiang, W. Li, L. Xie, and E. Yin, ‘‘Data
augmentation: Using channel-level recombination to improve classifica-
posed method, namely ShallowConvNet, DeepConvNet, and tion performance for motor imagery EEG,’’ Frontiers Hum. Neurosci.,
EEGNet. For each classifier, a two-step DA technique was vol. 15, Mar. 2021, Art. no. 645952.
applied: i) a percentage of the test data was considered avail- [8] Y. Pei, T. Sheng, Z. Luo, L. Xie, W. Li, Y. Yan, and E. Yin, ‘‘A tensor-based
frequency features combination method for brain–computer interfaces,’’ in
able as a calibration set during the training stage, in order Cognitive Systems and Information Processing. Suzhou, China: Springer,
to standardize data per subject, and ii) a similarity mea- Nov. 2021, pp. 511–526.
sure between subjects was considered in the creation of the [9] P. Arpaia, A. Esposito, A. Natalizio, and M. Parvis, ‘‘How to successfully
classify EEG in motor imagery BCI: A metrological analysis of the state
validation sets. This was made with the aim to reduce the of the art,’’ J. Neural Eng., vol. 19, no. 3, Jun. 2022, Art. no. 031002.
high non-stationarity typical of the brain signals, leading to [10] P. Arpaia, D. Coyle, F. Donnarumma, A. Esposito, A. Natalizio, and
improved classification accuracy. The experimental results M. Parvis, ‘‘Visual and haptic feedback in detecting motor imagery within
a wearable brain–computer interface,’’ Measurement, vol. 206, Jan. 2023,
were obtained by testing the proposal on a benchmark dataset, Art. no. 112304.
composed of 35 subjects and 40 simultaneous flickering [11] P. Arpaia, C. Bravaccio, G. Corrado, L. Duraccio, N. Moccaldi, and
stimuli. A 10-run Hold-Out Validation was used to simu- S. Rossi, ‘‘Robotic autism rehabilitation by wearable brain-computer inter-
face and augmented reality,’’ in Proc. IEEE Int. Symp. Med. Meas. Appl.
late real-world usage in a statistically significant way. The
(MeMeA), Jun. 2020, pp. 1–6.
experimental results show that the proposed DA approach [12] P. Arpaia, E. De Benedetto, N. Donato, L. Duraccio, and N. Moccaldi,
significantly helps to improve classification accuracy both ‘‘A wearable SSVEP BCI for AR-based, real-time monitoring applica-
over standard ML strategies and FBCCA, which is currently tions,’’ in Proc. IEEE Int. Symp. Med. Meas. Appl. (MeMeA), Jun. 2021,
pp. 1–6.
considered the gold standard in terms of SSVEPs classifi- [13] P. Arpaia, S. Criscuolo, E. De Benedetto, N. Donato, and L. Duraccio,
cation [27]. In fact, the low p-values obtained suggest that, ‘‘A wearable AR-based BCI for robot control in ADHD treatment: Pre-
for short-time signals, a Dataset Shift problem may arise liminary evaluation of adherence to therapy,’’ in Proc. 15th Int. Conf. Adv.
Technol., Syst. Services Telecommun. (TELSIKS), Oct. 2021, pp. 321–324.
due to the non-stationarity of EEG signals. Performance of [14] C.-M. Wu, Y.-J. Chen, I. A. E. Zaeni, and S.-C. Chen, ‘‘A new SSVEP
62.27 % accuracy was achieved by EEGNet - DA with a based BCI application on the mobile robot in a maze game,’’ in Proc. Int.
time window of only 1.0 s, while EEGNet - Standard and Conf. Adv. Mater. Sci. Eng. (ICAMSE), Nov. 2016, pp. 550–553.
[15] Y. Li and T. Kesavadas, ‘‘SSVEP-based brain-computer interface for part-
FBCCA reached 56.82 % and 39.34 %, respectively. This picking robotic co-worker,’’ J. Comput. Inf. Sci. Eng., vol. 22, no. 2,
allows easier development of SSVEP-based BCIs in contexts Apr. 2022, Art. no. 021001.
where short-time signals are required for real-time human- [16] L. Chen, P. Chen, S. Zhao, Z. Luo, W. Chen, Y. Pei, H. Zhao, J. Jiang,
M. Xu, Y. Yan, and E. Yin, ‘‘Adaptive asynchronous control system of
environment interactions. Future work will be dedicated to
robotic arm based on augmented reality-assisted brain–computer inter-
further enhance the classification performance in order to face,’’ J. Neural Eng., vol. 18, no. 6, Dec. 2021, Art. no. 066005.
achieve satisfactory results even with time windows equal to [17] B. Liu, X. Huang, Y. Wang, X. Chen, and X. Gao, ‘‘BETA: A large
or lower than 1.0 s. benchmark database toward SSVEP-BCI application,’’ Frontiers Neu-
rosci., vol. 14, p. 627, Jun. 2020.
[18] Q. Wei, M. Xiao, and Z. Lu, ‘‘A comparative study of canonical correlation
REFERENCES analysis and power spectral density analysis for SSVEP detection,’’ in
[1] P. Arpaia, L. Duraccio, N. Moccaldi, and S. Rossi, ‘‘Wearable brain– Proc. 3rd Int. Conf. Intell. Hum.-Mach. Syst. Cybern., Aug. 2011, pp. 7–10.
computer interface instrumentation for robot-based rehabilitation by [19] S. Zhang, X. Gao, and X. Chen, ‘‘Humanoid robot walking in maze
augmented reality,’’ IEEE Trans. Instrum. Meas., vol. 69, no. 9, controlled by SSVEP-BCI based on augmented reality stimulus,’’ Frontiers
pp. 6362–6371, Sep. 2020. Hum. Neurosci., vol. 16, p. 16, Jul. 2022.
[2] J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller, and [20] Y. Ke, P. Liu, X. An, X. Song, and D. Ming, ‘‘An online SSVEP-BCI system
T. M. Vaughan, ‘‘Brain–computer interfaces for communication and con- in an optical see-through augmented reality environment,’’ J. Neural Eng.,
trol,’’ Clin. Neurophysiol., vol. 113, no. 6, pp. 767–791, 2002. vol. 17, no. 1, Feb. 2020, Art. no. 016066.

VOLUME 11, 2023 36155

 A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

[21] M. Azeem, A. Haleem, and M. Javaid, ‘‘Symbiotic relationship between [42] C. Jia, X. Gao, B. Hong, and S. Gao, ‘‘Frequency and phase mixed coding
machine learning and industry 4.0: A review,’’ J. Ind. Integr. Manag., vol. 7, in SSVEP-based brain–computer interface,’’ IEEE Trans. Biomed. Eng.,
no. 3, pp. 401–433, 2022. vol. 58, no. 1, pp. 200–206, Jan. 2011.
[22] K.-R. Müller, M. Krauledat, G. Dornhege, G. Curio, and B. Blankertz, [43] G. R. Müller-Putz, R. Scherer, C. Brauneis, and G. Pfurtscheller, ‘‘Steady-
‘‘Machine learning techniques for brain-computer interfaces,’’ Biomed. state visual evoked potential (SSVEP)-based communication: Impact of
Tech, vol. 49, no. 1, pp. 11–22, Dec. 2004. harmonic frequency components,’’ J. Neural Eng., vol. 2, no. 4, p. 123,
[23] R. Singla and B. Haseena, ‘‘Comparison of SSVEP signal classification 2005.
techniques using SVM and ANN models for BCI applications,’’ Int. J. Inf. [44] G. Hakvoort, B. Reuderink, and M. Obbink, ‘‘Comparison of PSDA and
Electron. Eng., vol. 4, no. 1, p. 6, 2014. CCA detection methods in a SSVEP-based BCI-system,’’ Univ. Twente,
[24] M. Farooq and O. Dehzangi, ‘‘High accuracy wearable SSVEP detec- Centre Telematics Inf. Technol., Bengaluru, Karnataka, Tech. Rep. TR-
tion using feature profiling and dimensionality reduction,’’ in Proc. IEEE CTIT-11-03, 2011.
14th Int. Conf. Wearable Implant. Body Sensor Netw. (BSN), May 2017, [45] X. Chen, Y. Wang, M. Nakanishi, X. Gao, T.-P. Jung, and S. Gao, ‘‘High-
pp. 161–164. speed spelling with a noninvasive brain–computer interface,’’ Proc. Nat.
Acad. Sci. USA, vol. 112, no. 44, pp. E6058–E6067, Nov. 2015.
[25] I. A. Ansari, R. Singla, and M. Singh, ‘‘SSVEP and ANN based optimal
[46] Z. Lin, C. Zhang, W. Wu, and X. Gao, ‘‘Frequency recognition based
speller design for brain computer interface,’’ Comput. Sci. Techn., vol. 2,
on canonical correlation analysis for SSVEP-based BCIs,’’ IEEE Trans.
no. 2, pp. 338–349, Jul. 2015.
Biomed. Eng., vol. 53, no. 12, pp. 2610–2614, Dec. 2006.
[26] T.-H. Nguyen and W.-Y. Chung, ‘‘A single-channel SSVEP-based BCI
[47] P. Arpaia, E. De Benedetto, L. De Paolis, G. D’Errico, N. Donato,
speller using deep learning,’’ IEEE Access, vol. 7, pp. 1752–1763,
and L. Duraccio, ‘‘Performance enhancement of wearable instrumenta-
2019.
tion for AR-based SSVEP BCI,’’ Measurement, vol. 196, Jun. 2022,
[27] J. J. Podmore, T. P. Breckon, N. K. Aznan, and J. D. Connolly, ‘‘On the rel- Art. no. 111188.
ative contribution of deep convolutional neural networks for SSVEP-based [48] P. Arpaia, E. De Benedetto, L. De Paolis, G. D’Errico, N. Donato, and
bio-signal decoding in BCI speller applications,’’ IEEE Trans. Neural Syst. L. Duraccio, ‘‘Highly wearable SSVEP-based BCI: Performance compar-
Rehabil. Eng., vol. 27, no. 4, pp. 611–618, Apr. 2019. ison of augmented reality solutions for the flickering stimuli rendering,’’
[28] X. Chen, Y. Wang, S. Gao, T.-P. Jung, and X. Gao, ‘‘Filter bank canon- Meas., Sensors, vol. 18, Dec. 2021, Art. no. 100305.
ical correlation analysis for implementing a high-speed SSVEP-based [49] Y. Chen, C. Yang, X. Chen, Y. Wang, and X. Gao, ‘‘A novel training-
brain–computer interface,’’ J. Neural Eng., vol. 12, no. 4, Aug. 2015, free recognition method for SSVEP-based BCIs using dynamic window
Art. no. 046008. strategy,’’ J. Neural Eng., vol. 18, no. 3, Jun. 2021, Art. no. 036007.
[29] D. Huang, S. Zhou, and D. Jiang, ‘‘Generator-based domain adaptation [50] A. Apicella, P. Arpaia, E. De Benedetto, N. Donato, L. Duraccio,
method with knowledge free for cross-subject EEG emotion recognition,’’ S. Giugliano, and R. Prevete, ‘‘Enhancement of SSVEPs classifica-
Cogn. Comput., vol. 14, pp. 1316–1327, Apr. 2022. tion in BCI-based wearable instrumentation through machine learn-
[30] D. Ibáñez-Soria, A. Soria-Frisch, J. Garcia-Ojalvo, and G. Ruffini, ‘‘Char- ing techniques,’’ IEEE Sensors J., vol. 22, no. 9, pp. 9087–9094,
acterization of the non-stationary nature of steady-state visual evoked May 2022.
potentials using echo state networks,’’ PLoS ONE, vol. 14, no. 7, Jul. 2019, [51] X. Glorot and Y. Bengio, ‘‘Understanding the difficulty of training deep
Art. no. e0218771. feedforward neural networks,’’ in Proc. 13th Int. Conf. Artif. Intell. Statist.,
[31] D. Kapgate, D. Kalbande, and U. Shrawankar, ‘‘Adaptive classifi- 2010, pp. 249–256.
cation to reduce non-stationarity in visual evoked potential brain- [52] D. P. Kingma and J. Ba, ‘‘Adam: A method for stochastic optimization,’’
computer interfaces,’’ Bio-Algorithms Med-Syst., vol. 15, no. 2, Jun. 2019, 2014, arXiv:1412.6980.
Art. no. 20190020. [53] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, ‘‘Fast and accurate
[32] J. Quinonero-Candela, M. Sugiyama, A. Schwaighofer, and deep network learning by exponential linear units (ELUs),’’ 2015,
N. D. Lawrence, Dataset Shift in Machine Learning. Cambridge, arXiv:1511.07289.
MA, USA: MIT Press, 2008. [54] S. Kullback and R. A. Leibler, ‘‘On information and sufficiency,’’ Ann.
[33] J. L. Hagad, T. Kimura, K.-I. Fukui, and M. Numao, ‘‘Learning subject- Math. Statist., vol. 22, no. 1, pp. 79–86, 1951.
generalized topographical EEG embeddings using deep variational autoen- [55] L. Angrisani, P. Arpaia, A. Esposito, and N. Moccaldi, ‘‘A wearable brain–
coders and domain-adversarial regularization,’’ Sensors, vol. 21, no. 5, computer interface instrument for augmented reality-based inspection in
p. 1792, Mar. 2021. industry 4.0,’’ IEEE Trans. Instrum. Meas., vol. 69, no. 4, pp. 1530–1539,
[34] Y. Wang, X. Chen, X. Gao, and S. Gao, ‘‘A benchmark dataset for SSVEP- Apr. 2020.
based brain–computer interfaces,’’ IEEE Trans. Neural Syst. Rehabil. Eng., [56] X. Zhao, Y. Du, and R. Zhang, ‘‘A CNN-based multi-target fast classifi-
vol. 25, no. 10, pp. 1746–1752, Oct. 2017. cation method for AR-SSVEP,’’ Comput. Biol. Med., vol. 141, Feb. 2022,
[35] C. M. Bishop and N. M. Nasrabadi, Pattern Recognition and Machine Art. no. 105042.
Learning, vol. 4. Berlin, Germany: Springer-Verlag, 2006. [57] S. G. K. Patro and K. K. Sahu, ‘‘Normalization: A preprocessing stage,’’
2015, arXiv:1503.06462.
[36] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter,
[58] J. Bergstra and Y. Bengio, ‘‘Random search for hyper-parameter
K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball,
optimization,’’ J. Mach. Learn. Res., vol. 13, no. 2, pp. 281–305,
‘‘Deep learning with convolutional neural networks for EEG decoding
2012.
and visualization,’’ Hum. Brain Mapping, vol. 38, no. 11, pp. 5391–5420,
[59] D. Kingma and J. Ba, ‘‘Adam: A method for stochastic optimization,’’ in
Aug. 2017.
Proc. Int. Conf. Learn. Represent., Dec. 2014, pp. 1–15.
[37] V. Lawhern, A. Solon, N. Waytowich, S. M. Gordon, C. Hung, and
B. J. Lance, ‘‘EEGNet: A compact convolutional neural network for EEG-
based brain–computer interfaces,’’ J. Neural Eng., vol. 15, no. 5, p. 056013,
2018.
[38] W. Zhou, A. Liu, and X. Chen, ‘‘Compact CNN with dynamic window for
SSVEP-based BCIs,’’ in Proc. 41st Chin. Control Conf. (CCC), Jul. 2022,
pp. 3097–3101.
[39] N. Waytowich, V. J. Lawhern, J. O. Garcia, J. Cummings, J. Faller, P. Sajda,
and J. M. Vettel, ‘‘Compact convolutional neural networks for classifi-
cation of asynchronous steady-state visual evoked potentials,’’ J. Neural
Eng., vol. 15, no. 6, Dec. 2018, Art. no. 066031. ANDREA APICELLA received the M.S. degree in computer science and the
[40] L. Carretié, ‘‘Exogenous (automatic) attention to emotional stimuli: Ph.D. degree in mathematics and computer science from the University of
A review,’’ Cogn., Affect., Behav. Neurosci., vol. 14, no. 4, pp. 1228–1258, Naples Federico II, Italy, in 2014 and 2019, respectively. He is currently a
2014. Researcher with the Department of Information Technology and Electrical
[41] Y. Wang, R. Wang, X. Gao, B. Hong, and S. Gao, ‘‘A practical VEP-based Engineering, University of Naples Federico II. His current research interests
brain-computer interface,’’ IEEE Trans. Neural Syst. Rehabil. Eng., vol. 14, include artificial intelligence methods and eXplainable Artificial Intelligence
no. 2, pp. 234–240, Jun. 2006. (XAI) approaches for explaining the AI system’s decisions.

36156 VOLUME 11, 2023

A. Apicella et al.: Employment of DA Techniques in SSVEP-Based Brain–Computer Interfaces

PASQUALE ARPAIA (Senior Member, IEEE) received the M.S. and Ph.D. LUIGI DURACCIO received the M.S. degree (cum laude) in electronic
degrees in electrical engineering from the University of Naples Federico engineering from the University of Naples Federico II, in 2018. He developed
II, Naples, Italy, in 1987 and 1992, respectively. He was an Associate his master’s thesis with CERN, Geneva, Switzerland, in the field of radiation
Professor with the University of Sannio, Benevento, Italy. He is currently measurement for electronics. His current research interests include biomed-
a Full Professor of instrumentation and measurements with the University of ical instrumentation and measurement, electroencephalographic data acqui-
Naples Federico II, and a Team Leader with CERN, Geneva, Switzerland. sition and processing, augmented reality, and brain–computer interfaces.
He is also the Head of the Interdepartmental Research Centre in Health
Management and Innovation in Healthcare, University of Naples Federico
II. His current research interests include augmented reality, brain–computer
interfaces, cyber-security, digital instrumentation and measurement tech-
niques, evolutionary diagnostics, and distributed measurement systems. SALVATORE GIUGLIANO received the M.S. degree (cum laude) in com-
puter science from the University of Naples Federico II, Naples, Italy,
in 2019. He developed his master’s thesis in neural networks. He currently
EGIDIO DE BENEDETTO (Senior Member, IEEE) received the M.S. degree collaborated as a Consultant and a Researcher with Villa delle Ginestre
in materials engineering and the Ph.D. degree in information engineering Clinic, Volla, Italy. His current research interests include analysis and inter-
from the University of Salento, Lecce, Italy, in 2006 and 2010, respectively. pretation of EEG signals with machine learning techniques, transfer learning
He was with the Institute of Microelectronics and Microsystems, National on EEG data, and eXplainable artificial intelligence. In 2023, he was awarded
Research Council, Lecce, from 2010 to 2012. From 2012 to 2019, he was with the Ph.D. degree from the University of Naples Federico II.
a Research Fellow with the University of Salento. Since 2019, he has been
an Associate Professor with the Department of Electrical Engineering and
Information Technology, University of Naples Federico II, Italy.
ROBERTO PREVETE received the M.Sc. degree in physics and the Ph.D.
NICOLA DONATO (Senior Member, IEEE) received the M.S. degree in degree in mathematics and computer science from the Department of Elec-
electronic engineering from the University of Messina, Messina, Italy, and trical Engineering and Information Technologies (DIETI), University of
the Ph.D. degree from the University of Palermo, Palermo, Italy. He is Naples Federico II, Naples, Italy. He is currently an Assistant Professor
currently an Associate Professor of electrical and electronic measurements of computer science with DIETI, University of Naples Federico II. His
and the Head of the Laboratories of Electronics for Sensors and for Systems research has published in international journals, such as Biological Cyber-
of Transduction and Electrical and Electronic Measurements, University of netics, Experimental Brain Research, Neurocomputing, Neural Networks,
Messina. His current research interests include sensor characterization and and Behavioral and Brain Sciences. His current research interests include
modeling, the development of measurement systems for sensors, and the computational models of brain mechanisms, machine learning, and artificial
characterization of electronic devices up to microwave range and down to neural networks and their applications.
cryogenic temperatures.

Open Access funding provided by ‘Università degli Studi di Napoli "Federico II"’ within the CRUI CARE Agreement

VOLUME 11, 2023 36157