TYPE Original Research

PUBLISHED 04 October 2023

DOI 10.3389/fnins.2023.1246940

A novel untrained SSVEP-EEG

OPEN ACCESS feature enhancement method
using canonical correlation
EDITED BY
Vince D. Calhoun,
Georgia State University, United States

REVIEWED BY
Rui Li,
analysis and underdamped
Xi'an University of Technology, China
Qingshan She,
Hangzhou Dianzi University, China
second-order stochastic
*CORRESPONDENCE
Guanghua Xu
resonance
ghxu@mail.xjtu.edu.cn

RECEIVED 29June 2023 Ruiquan Chen 1, Guanghua Xu 2*, Huanqing Zhang 1, Xun Zhang 1,
ACCEPTED 19 September 2023
PUBLISHED 04 October 2023
Baoyu Li 1, Jiahuan Wang 1 and Sicong Zhang 1
CITATION School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an, China, 2 State Key Laboratory for
1

Chen R, Xu G, Zhang H, Zhang X, Li B, Manufacturing Systems Engineering, School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an,
Wang J and Zhang S (2023) A novel untrained China
SSVEP-EEG feature enhancement method
using canonical correlation analysis and
underdamped second-order stochastic
resonance.
Front. Neurosci. 17:1246940. Objective: Compared with the light-flashing paradigm, the ring-shaped motion
doi: 10.3389/fnins.2023.1246940 checkerboard patterns avoid uncomfortable flicker or brightness modulation,
COPYRIGHT improving the practical interactivity of brain-computer interface (BCI) applications.
© 2023 Chen, Xu, Zhang, Zhang, Li, Wang and However, due to fewer harmonic responses and more concentrated frequency
Zhang. This is an open-access article
energy elicited by the ring-shaped checkerboard patterns, the mainstream
distributed under the terms of the Creative
Commons Attribution License (CC BY). The untrained algorithms such as canonical correlation analysis (CCA) and filter
use, distribution or reproduction in other bank canonical correlation analysis (FBCCA) methods have poor recognition
forums is permitted, provided the original
performance and low information transmission rate (ITR).
author(s) and the copyright owner(s) are
credited and that the original publication in this Methods: To address this issue, a novel untrained SSVEP-EEG feature enhancement
journal is cited, in accordance with accepted
method using CCA and underdamped second-order stochastic resonance (USSR)
academic practice. No use, distribution or
reproduction is permitted which does not is proposed to extract electroencephalogram (EEG) features.
comply with these terms.
Results: In contrast to typical unsupervised dimensionality reduction methods
such as common average reference (CAR), principal component analysis (PCA),
multidimensional scaling (MDS), and locally linear embedding (LLE), CCA exhibits
higher adaptability for SSVEP rhythm components.
Conclusion: This study recruits 42 subjects to evaluate the proposed method
and experimental results show that the untrained method can achieve higher
detection accuracy and robustness.
Significance: This untrained method provides the possibility of applying a
nonlinear model from one-dimensional signals to multi-dimensional signals.

KEYWORDS

motion checkerboard patterns, brain-computer interface, canonical correlation

analysis, underdamped second-order stochastic resonance, information transmission
rate

Frontiers in Neuroscience 01 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

1. Introduction et al., 2014b), likelihood ratio test (Zhang et al., 2014c), and stochastic
resonance analysis (Chen et al., 2021a), have also proven to have
Brain-computer interface (BCI) is a normal output pathway unique advantages in SSVEP recognition.
system that does not rely on the composition of peripheral nerves and Benzi et al. (1981) first proposed the concept of stochastic
muscles, and can directly convert central nervous activities into resonance (SR) when studying the problem of global glacial periods,
artificial output (Vidal, 1973; Li et al., 2017; Jin et al., 2021). and successfully applied it to explain the phenomenon of periodic
Steady-state visual evoked potentials (SSVEPs) based on BCI changes in paleoclimate. Shortly after, Fauve and Heslot (1983)
systems have the advantages of short training time, high signal-to- observed the phenomenon of SR while studying the synchronization
noise ratio, and short response time, and are widely used in clinical of noise-induced transitions in a bistable system experiment with a
detection technology (Li et al., 2022). When an external visual trigger circuit. Mcnamara et al. (1988) once again verified the
stimulus of constant frequencies is applied, the neural network existence of bistable stochastic resonance (BSR) in the ring laser
consistent with the stimulation frequency or harmonic components experiment. Collins et al. (1995) first extended SR theory to the field
will generate resonance, causing the brain’s potential activity to change of aperiodic signal processing when studying FitzHugh Nagumo
significantly at the stimulation frequency or harmonic components, (FHN) neuron models. Recently, Chen et al. (2021a,b) and Chen et al.
resulting in SSVEP signals. SSVEP signals can exhibit spectral peaks (2022) demonstrated in experiments that the FHN neuron model can
at stimulation frequency or harmonic components in the power effectively enhance the feature responses of EEG signals, thereby
spectrum of EEG signals (Kramer et al., 2021). By analyzing and improving recognition accuracy, regardless of the time domain,
detecting the frequency corresponding to the spectral peak, it is frequency domain, or time-frequency domain. Lu et al. (2015)
possible to detect the stimulus source of the subject’s visual gaze, proposed underdamped second-order stochastic resonance (USSR)
thereby identifying the subject’s intention. However, although SSVEP to improve weak signal detection technology. This novel model
induced by the motion checkerboard paradigm can reduce visual considers the system inertia and underdamped damping factor based
fatigue in subjects, due to its generation mechanism, there are almost on bistable stochastic resonance (BSR), which is more conducive to
no harmonic components and frequency energy is more concentrative, high SNR output. Traditional denoising methods improve the SNR
thus leading to low recognition accuracy (Han et al., 2018). by suppressing noise, which may result in the loss of useful features.
The first application in SSVEP feature classification is the Power However, SR utilizes the synergistic effects of input signals, noise, and
Spectral Density Analysis (PSDA) algorithm (Ming Cheng et al., resonance systems to enhance feature responses of signals, and has
2002), which uses the Fast Fourier Transform (FFT) to convert SSVEP excellent nonlinear signal detection capabilities with noise immunity.
signals from the time domain to the frequency domain, thereby The main contribution of this study is to propose a novel
obtaining the amplitude and phase characteristics of each stimulus untrained SSVEP feature enhancement method using CCA
frequency. Because this method only analyzes one electrode signal in dimensionality reduction technology and the USSR model. Compared
multi-channel signals, the obtained signal has a low signal-to-noise with mainstream unsupervised dimensionality reduction methods,
ratio (SNR). Wavelet Transform (WT) can be regarded as a Fourier such as common average reference (CAR; Orekhova et al., 2002),
Transform with an adjustable window, which provides both relevant principal component analysis (PCA; Wold et al., 1987),
frequency components and occurrence time information, but it still multidimensional scaling (MDS; Saeed et al., 2018), and locally linear
cannot identify nonlinear signals well (Hu et al., 2014). Volosyak et al. embedding (LLE; De Ridder et al., 2003), the CCA method reflects a
proposed the Minimum Energy Combination (MEC) algorithm high degree of matching with SSVEP signals. Experimental results
(Friman et al., 2007), which mainly seeks a spatial filter to project show that the CCA-USSR method has higher recognition accuracy,
multi-channel signals into low dimensional space. MEC can effectively ITR, and better robustness in all subjects.
reduce background noise, but useful information in EEG signals may The rest of this article is arranged as follows: section 2 introduces
be lost in the linear transformation. The canonical correlation analysis in detail typical dimensionality reduction methods and the novel
(CCA) algorithm was first applied to SSVEP classification by Lin et al. CCA-USSR framework proposed in this study. In section 3, the
(2006). CCA mainly projects multi-channel SSVEP signals and specific experiments and results obtained by different methods are
corresponding reference signals into a low dimensional space through explained. Compared with the CAR, PCA, MDS, and LLE methods,
a spatial filter and then calculates the correlation between the two. The the CCA-USSR method showed a better BCI performance in all
maximum value of the correlation coefficient corresponds to the subjects. The processing results of each method are discussed in
stimulation frequency, which is superior to the MEC algorithm. section 4. Finally, section 5 provides the conclusions.
Currently, many variants of the CCA method have achieved excellent
BCI performance, such as multi-way canonical correlation analysis
(MwayCCA; Zhang et al., 2011), multi-set canonical correlation 2. Methodology
analysis (MestCCA; Zhang et al., 2014a), Filter bank canonical
correlation analysis (FBCCA; Chen and Gao, 2015), Task-related 2.1. The USSR model and standard FBCCA
component analysis (TRCA; Nakanishi et al., 2017), Task-discriminant method
component analysis (TDCA; Liu et al., 2021) and so on. Among them,
the FBCCA method is the most effective and widely used untrained 2.1.1. USSR
method in SSVEP-EEG detection technology. Nevertheless, due to the The Langevin equation for the BSR model is an overdamped
low harmonic components of SSVEP induced by motion checkerboard first-order differential equation due to neglecting the inertia term
patterns, the FBCCA method cannot play its role. In addition, some and normalizing the damping factor. Nevertheless, it has been
novel methods, such as multivariate synchronization index (Zhang proven (Lu et al., 2015) that the system inertia and damping factor

Frontiers in Neuroscience 02 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

can facilitate high-SNR output. Considering these two factors, the high-amplitude artifacts at the selected reference electrode. Therefore,
BSR model is improved into a second-order differential equation the selection of reference electrodes is crucial for the CAR method.
which is called the USSR model. Hence, the USSR model can The single-electrode output potential VCAR between the electrode i and
be expressed as the reference electrode can be expressed as

d2x dU ( x ) dx dx 1 n
dt 2
=−
dx
−β
dt
= ax − bx 3 − β
dt
+ s ( t ) + n ( t ) (1) VCAR = Vreference − ∑Vi
n i =1
(4)

where a and b are the system parameters satisfying a, b ∈ R +, x is Where Vreference is the selected reference signal, n is the number
the output signal, s ( t ) is the input signal. n ( t ) is the Gaussian white of electrodes.
noise with a mean value of zero and an autocorrelation function
satisfying n ( t1 ) n ( t2 ) = 2 Dδ ( t1 − t2 ) (Wu and Zhu, 2008); . represents 2.2.2. PCA
the overall mean value; 0 < β < 1 is the damping factor. The fourth- It is one of the most popular unsupervised linear dimensionality
order Runge-Kuta algorithm with fixed steps is used to solve the reduction methods nowadays. Its main idea is to obtain a new matrix
differential equations with higher solution accuracy. To meet the with the largest variance in the projected dimension after data is
needs of differential equation calculation, the input signal needs to multiplied by a matrix, thereby using fewer data dimensions while
be transformed into a one-dimensional vector. retaining the characteristics of more original data. The measure of
information quantity is the variance of data, which is described
2.1.2. FBCCA as follows
As a standard untrained SSVEP recognition algorithm, the
m
FBCCA method uses CCA to calculate the canonical correlation
Var ( yi ) = aiT ∑ai (5)
coefficient ρi of each sub-band signal ( X i ,i = 1,2,…,N ) which is i =1
divided via multiple filter banks. The feature discrimination coefficient
at the i-th target frequency is obtained by
where ai is the i-th transformation vector, Σ is the covariance
n matrix of the original data, yi is the i-th principal component, Var ( yi )
ρ = ∑w ( i ) ⋅ ( ρi )
2
(2) is the variance of the i-th principal component.
i =1
The PCA method uses orthogonal transformations to convert the
observed data into principal components represented by linear
where w ( i ) is the weight of the i-th sub-band signal which can independent variables, the number of which is usually smaller than
be obtained by the number of original variables. Thus, PCA is a common
dimensionality reduction method using the linear projection rule.
w ( i ) = i − a + b, i ∈ [1 N ] (3)
2.2.3. MDS
It uses the paired similarity of samples to construct a
As previously reported (Chen et al., 2015), a and b constants are low-dimensional space so that the distance of each pair of samples in
set to 1.25 and 0.25. the high-dimensional space is as consistent as possible with the sample
similarity in the constructed low-dimensional space. A greater
similarity between two objects can be reflected by a smaller distance
2.2. Typical unsupervised dimensionality in MDS space. The basic principle is described by
reduction methods
p
∑ ( yij − yik )
2
Data dimensionality reduction can be used as a means of feature δ jk = (6)
i =1
extraction: to identify the main features from the original features of
the dataset, that is, the features that best describe the distribution of
data in the dataset. In other words, while preserving the main features where δ jk is the dissimilarity between samples j and k, p is the
of the dataset, high-dimensional data is projected into a number of properties used to perform MDS, and y is the elemental
low-dimensional feature space. Since the input requirement of the concentration or index in this study.
USSR model is a one-dimensional vector, a data dimensionality Hence, classic MDS performs dimensionality reduction on high-
reduction method that matches the EEG rhythm features is required. dimensional data while ensuring a consistent distance between the
The following describes five typical unsupervised dimensionality original space and low-dimensional spatial samples.
reduction methods.
2.2.4. LLE
2.2.1. CAR It is one of the commonly used manifold learning methods and a
The principle is to calculate the average signal of all recording nonlinear dimensionality reduction method suitable for processing
electrodes, and then subtract this average value from the selected nonlinear data. It is based on the manifold assumption that data in
reference electrode. However, the method is influenced strongly by high-dimensional space is distributed on low-dimensional manifolds.

Frontiers in Neuroscience 03 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

The LLE dimensionality reduction method can be described as the  cos ( 2π f d t ) 

 
following three steps:  sin ( 2π f d t ) 
 
Y ( fd ) =  …  , t = 1 / f s ,…, N / f s (11)
1. For each data point xi, find its K nearest neighbors. cos ( 2kπ f t ) 
 d 
2. Compute the reconstruction weights of the neighbors that
minimize the error of reconstructing xi.  sin ( 2kπ f d t ) 

where k is the number of harmonics of SSVEP signals; f d is the

N N
stimulation frequency of the template signals; N is the number of
ε (W ) = ∑xi − ∑wij x j 2
(7)
i =1 j =1
sample points; f s is the sampling frequency.

N
{ }
Subject to wij = 0, if x j ∉ x1i ,xi2 ,…,xiK , and ∑wij = 1. 2.3. Feature enhancement for SSVEP using
j =1 canonical correlation analysis and
3. Compute the low-dimensional embedding Y for yi that best underdamped second-order stochastic
preserves the local geometry represented by the resonance
reconstruction weights.
When the human eye receives a fixed frequency of visual
N N
stimulation, the potential activity of the cerebral cortex will
δ (Y ) = ∑ yi − ∑wij y j 2
(8)
i =1 j =1
be modulated to produce a continuous response related to the
stimulation frequency. This response has a periodic rhythm similar to
N N the visual stimulation, that is, the steady-state visual evoked potential.
Subject to 1 / N ∑ yi yiT = I , and ∑ yi = 0, where 0 is a column The SSVEP signal can exhibit spectral peaks at the stimulus frequency
i =1 i =1 or harmonic components in the power spectrum. By analyzing the
vector of zeros and I is an identity matrix. By the Rayleigh-Ritz frequency corresponding to the spectral peak, the stimulus source of
theorem (Luce and Perry, 1949), minimizing (10) with respect to the the subject’s visual gaze can be detected, thereby identifying the subject’s
yi’s can be done via finding the eigenvectors with the smallest intention. The novel untrained framework based on the CCA
(nonzero) eigenvalues. dimensionality reduction method and USSR model proposed in this
study is shown in Figure 1. Detailed procedures are described as follows:
2.2.5. CCA
The projection principle selected by the CCA method is that after a. Signal acquisition and preprocessing. Since the raw SSVEP
dimensionality reduction, the correlation coefficient of the two sets of signals are usually weak and mixed with multi-scale noise, it is
data is the largest. For input EEG data X and the reference signal Y , difficult to extract gaze frequencies in a single trial. Some
the goal of CCA is to find weight vectors w x and w y , so that the preprocessing steps such as filtering techniques need to be used
one-dimensional vectors obtained after X and Y projection are X ′ to remove noise interference. As previously described (Han
and Y ′, respectively. Therefore, one-dimensional vectors are et al., 2018), since the motion checkerboard pattern has few
obtained by harmonic components, in our study, we only consider the
fundamental frequency and the primary harmonic to achieve
X ′ = w xT X the highest ITR. Hence, a Butterworth filter with a passband
range of 3–40 Hz is selected to remove noise and some high-
frequency components.
Y ′ = w yT Y (9) b. Dimensionality reduction. Although the SR model can
effectively enhance the feature frequency of SSVEP signals, due
to the characteristic of its differential equation, the input signal
The optimization goal of the CCA dimensionality reduction needs to be transformed into a one-dimensional vector.
method is to maximize ρ ( X ′,Y ′ ) to obtain the corresponding Common unsupervised dimensionality reduction methods
projection vectors w x and w y . Then the correlation coefficient ρ include CAR, PCA, MDS, LLE, and CCA. In our study, these
between weight vectors w x and w y was calculated by methods are compared to get the optimal dimensionality
reduction methods. In the CAR method, we choose the Oz
T
E  wTx XY ( f ) w y  channel as the reference channel. The experimental results
max w x w y , ρ ( X ′,Y ′ ) =   indicate that the Oz channel has the highest recognition
 wTy Y ( f )  accuracy compared with other channels. In the LLE method,
E  wTx XX T w x  E   (10)
  Y f T w  the number of nearest neighbor points is set to 40.
 ( ) y
c. SSVEP feature enhancement. Typical SR models include BSR,
FHN, and USSR, among which the USSR model has the best
The frequency corresponding to the maximum correlation BCI performance despite having the most parameters (Chen
coefficient is regarded as the gaze target of the subjects. et al., 2023). Therefore, in this study, the USSR model was used
The template signals of the CCA method are given by as a means of feature enhancement.

Frontiers in Neuroscience 04 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

FIGURE 1
The flowchart of SSVEP detection using the CCA dimensionality reduction method and USSR model.

visual stimuli, and 0.7 s of rest time. The experimental conditions can
be found in the study (Chen et al., 2021a). The experimental process
is displayed in Figure 3.

3.2. CCA coefficient spectrums analysis for

SSVEP

Based on the above analysis, we know that the SR model can utilize
the noise energy to enhance the target frequency through synergistic
effects and noise immunity. Therefore, the main research direction of
FIGURE 2 this paper is to preserve as many effective features as possible in the
The user interface of the 35 focused targets. original multi-channel signals. As representative dimensionality
reduction methods among the eight methods, we compared the
coefficient spectrums of CCA, PCA-USSR, LLE-USSR, and CCA-USSR
d. SSVEP feature recognition. Then, PSDA and CCA recognition in Figure 4. In this study, the representative subject (S29) with the 2 s
methods are used to identify the subjects’ gaze targets. data length was utilized to compare the BCI performance. The
e. Target discrimination. Finally, the recognition accuracy is correlation coefficient between the processed SSVEP signal and the
obtained by matching the recognition frequency with the reference signal at a 0.5 frequency interval of 1-40 Hz is described as
stimulus frequency. the CCA coefficient spectrum in our study.
The CCA coefficient spectrum between the filtered EEG and the
template is presented in Figure 4A. As a classic SSVEP recognition
3. Experiment and results method, the CCA method can effectively identify the target frequency
of most subjects. For example, the CCA coefficients corresponding to
3.1. Experiments and datasets the seven target frequencies of 3, 7, 3.5, 11.5, 6, 9, and 11 Hz (the CCA
spectrums marked by the green box) have the maximum amplitude in
The experiment data included 30 males and 12 females (42 the entire spectrum and the recognition succeeds. However, it is worth
subjects, average age ± SD, 27.2 ± 2.6) originating from Chen et al. noting that in these successful cases, the amplitudes of several
(2023). Each subject has a normal or corrected vision. interference peaks are very close to that of the target frequencies,
The ring-shaped checkerboards with radial contraction– making the recognition effect not ideal. On the other hand, the CCA
expansion motion were adopted as the visual stimuli in our coefficients corresponding to the five frequencies of 15, 16, 12, 19, and
experiment. The stimulus paradigm was arranged into a 5 × 7 matrix 14.5 Hz (the CCA spectrums marked by red boxes) have not the
with a horizontal and vertical separation of 100 pixels and 50 pixels maximum amplitude in the CCA spectrum, and their amplitudes are
between two adjacent stimuli, respectively. The frequency range of 35 second only to the maximum interference peak. Hence, the feature
focused targets was 3–20 Hz with a frequency interval of 0.5 Hz. The extraction for SSVEP recognition finally fails and there is a need for new
35 focused frequencies for each trial were presented simultaneously methods to enhance the energy of the target frequency and improve the
with the data sampled at 1,000 Hz, as shown in Figure 2. The BCI decoding performance under the motion checkerboard pattern.
g.USBamp (g.tec Inc., Austria) was utilized to record SSVEP signals According to the previous study (Yao et al., 2019), the optimal
and the channels were set according to the 10/20 electrode system. parameter combination for the USSR model is [a, b, 𝛽, h] = [0.1, 1,
These eight electrodes POz, PO3, PO4, PO5, PO6, Oz, O1, and O2 0.35, 0.1].
were used to record the raw SSVEP signal. SSVEP is a specific EEG Compared with the CCA method, the USSR-based methods reduce
signal generated by the occipital region of the brain. These eight the amplitude of interference peaks in the CCA coefficient spectrum
electrodes are located closest to the occipital lobe, so the signals and increase the energy of the target frequency, making BCI recognition
collected by them are less noisy and more stable. Each subject is more accurate. The amplitude of target frequencies depends on the
required to conduct 35 trials, each consisting of 0.3 s of cues, 3 s of matching between the dimensionality reduction method and the SSVEP

Frontiers in Neuroscience 05 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

rhythm components. As shown in Figure 4B, in the corresponding spectrums of 4.5, 7, 3.5, and 8 Hz (the coefficient spectrums marked in
coefficient spectrums at 17, 16, 12, 19, and 18.5 Hz (the coefficient red boxes), the PCA-USSR method cannot accurately identify target
spectrums marked in green boxes), which cannot be correctly identified frequencies. As shown in Figure 4C, the combination of the nonlinear
by the CCA method, the amplitudes of the focused frequencies are LLE method and the USSR model has similar results compared with the
enhanced by the PCA-USSR model and have exceeded the amplitude of linear dimensionality reduction method. For target frequencies that
the interference peaks. However, for the corresponding coefficient cannot be identified by the CCA method, some can be corrected by the
LLE-USSR method, while others cannot be identified.
As shown in Figure 4D, compared with the typical PCA-USSR and
LLE-USSR methods, the CCA-USSR method can significantly
increase the amplitude of the target frequency and has an energy
concentration effect that matches the checkerboard pattern. Therefore,
compared with the CCA method, the recognition accuracy of target
frequency has increased by approximately 43%. For example, for the
seven frequencies of 17, 9.5, 11.5, 10, 8.5, 19.5, and 4 Hz that cannot
be identified by the LLE-USSR method, the USSR model can further
enhance the energy of the target frequency and optimize the SSVEP
recognition performance. While, in the power spectrums
corresponding to 3, 3.5, 5, and 8 Hz, since the CCA dimensionality
FIGURE 3 reduction method does not retain the effective features of the original
The experimental process. SSVEP well, the USSR model incorrectly enhances the energy of other
frequencies, resulting in recognition failure.

A B
f =4.5H z f =13H z f =3Hz f =15H z f =18H z f =4.5H z f =13H z f =3Hz f =15H z f =18H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =17H z f =9.5H z f =7Hz f =15.5Hz f =3.5H z f =17H z f =9.5H z f =7Hz f =15.5Hz f =3.5H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =16H z f =12H z f =11.5Hz f =7.5H z f =6Hz f =16H z f =12H z f =11.5Hz f =7.5H z f =6Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =10H z f =19H z f =9Hz f =8.5H z f =6.5H z f =10H z f =19H z f =9Hz f =8.5H z f =6.5H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =10.5Hz f =18.5Hz f =11H z f =16.5Hz f =20H z f =10.5Hz f =18.5Hz f =11H z f =16.5Hz f =20H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =17.5Hz f =5.5H z f =19.5Hz f =14.5Hz f =4Hz f =17.5Hz f =5.5H z f =19.5Hz f =14.5Hz f =4Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =13.5Hz f =14H z f =5Hz f =8Hz f =12.5Hz f =13.5Hz f =14H z f =5Hz f =8Hz f =12.5Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40

C D
f =4.5Hz f =13Hz f =3Hz f =15Hz f =18Hz f =4.5H z f =13Hz f =3Hz f =15Hz f =18Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =17Hz f =9.5Hz f =7Hz f =15.5H z f =3.5Hz f =17Hz f =9.5H z f =7Hz f =15.5H z f =3.5H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =16Hz f =12Hz f =11.5H z f =7.5Hz f =6Hz f =16Hz f =12Hz f =11.5H z f =7.5H z f =6Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =10Hz f =19Hz f =9Hz f =8.5Hz f =6.5Hz f =10Hz f =19Hz f =9Hz f =8.5H z f =6.5H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =10.5H z f =18.5H z f =11Hz f =16.5H z f =20Hz f =10.5H z f =18.5H z f =11Hz f =16.5H z f =20Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =17.5H z f =5.5Hz f =19.5H z f =14.5H z f =4Hz f =17.5H z f =5.5H z f =19.5H z f =14.5H z f =4Hz
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40
f =13.5H z f =14Hz f =5Hz f =8Hz f =12.5H z f =13.5H z f =14Hz f =5Hz f =8Hz f =12.5H z
1 1 1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
0 0 0 0 0 0 0 0 0 0
0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40 0 20 40

FIGURE 4
(A) The coefficient spectrums of the CCA method. (B) The coefficient spectrums of the PCA-USSR method. (C) The coefficient spectrums of the LLE-
USSR method. (D) The coefficient spectrums of the CCA-USSR method.

Frontiers in Neuroscience 06 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

The above analysis has demonstrated that the CCA dimensionality The analysis data length of SSVEP signals is also an important
reduction method can preserve the effective features of the original indicator that significantly affects recognition accuracy and ITR. Here,
signal to the greatest extent and the nonlinear weak feature Figure 5 compares the average recognition accuracy and ITR of CCA,
enhancement based on USSR dynamics models is highly compatible FBCCA, and USSR-based methods under different data lengths (3,
with the non-stationary SSVEP. The USSR model takes advantage of 2.5, 2, 1.5, and 1 s).
the unique conversion of noise energy to signal energy, thereby From Figure 5, each method exhibits reliable results compared
enhancing the amplitude and the energy of focused targets and with the previous study (Yan et al., 2021), and among these methods,
improving the BCI decoding performance. Table 1 shows the average the CCA-USSR method has the highest recognition accuracy and ITR
accuracy and ITR of 42 subjects based on five dimensionality from 1 to 3 s data length. Note that due to the lack of harmonic
reduction methods and two standard methods for processing SSVEP components in the SSVEP signal induced by the motion checkerboard
signals (Data length T = 2 s). paradigm, the FBCCA method did not achieve ideal results in our
study. Especially, based on PSDA recognition methods, all USSR-
based methods can achieve the best BCI decoding performance only
3.3. BCI performance at 2 s date length. The experimental results show that the CCA-USSR
method can significantly outperform the classic CCA and FBCCA
Here, paired t-tests were performed to determine significant methods, as well as other USSR-based methods, at any data length
differences (defined as p < 0.05) in accuracy and ITR for different (p < 0.05). This also indicates that the CCA method has a high degree
methods. The information transfer rate (ITR) is an important and of matching with SSVEP, which can retain the most features of the
effective indicator to measure SSVEP-BCI recognition performance original multi-channel signal after dimensionality reduction. In
among different methods. It is used to express the amount of addition, the CAR-USSR method has excellent BCI performance
information transmitted in a unit of time. ITR can be obtained by comparable to the classic CCA method at 1 s data length. One possible
explanation is that CAR relies on electrode selection, so when the data
60   1 − σ  length is shorter, the CAR-USSR method can show better BCI
ITR = log 2 M + σ log 2 σ + (1 − σ ) log 2    (12)
T   M − 1  decoding ability. On the other hand, it is not difficult to infer that the
nonlinear LLE method is not suitable for non-stationary SSVEP
signals. One hypothesis is that multi-channel SSVEP signals do not
where σ refers to the average recognition accuracy, M refers to the meet the manifold distribution assumption of input signals. The
number of gaze frequencies and T refers to the data length for analysis. rationale we get from the experimental results is that although multi-
The higher the average recognition accuracy, the larger the number scale noise may have negative effects on SSVEP recognition, the CCA
of gaze targets, the shorter the used data length, and the higher the dimensionality reduction method can retain the most features of the
obtained ITR. Meanwhile, the data length also affects recognition original multi-channel SSVEP signal, and the USSR model can use
accuracy. For example, too short a data length may result in fewer noise energy to enhance the amplitude of the target frequency, thereby
recognizable SSVEP features and a decrease in recognition accuracy. improving the detection accuracy and ITR of SSVEP decoding.
Using the classic CCA method to identify the gaze frequencies of Besides, the processing time of unsupervised dimensionality
42 subjects can achieve an average accuracy of 61.16 ± 19.59 and an reduction methods was detected to further compare the online
ITR of 66.68 ± 33.48. As the state-of-the-art method for SSVEP calculation ability of different methods in Table 2.
recognition, the average accuracy and ITR of FBCCA are increased to From Table 2, we can see that the CAR and PCA methods have
65.99 ± 11.68 and 74.22 ± 20.66 bits/min, respectively. the shortest processing time and the CCA method comes next, while
It is worth noting that the CAR-USSR method is not only affected by the MDS and LLE methods have the slowest processing speed. In
the reference channel but also has the worst robustness among the five particular, the processing time of the LLE method is determined by
dimensionality reduction methods with a variance of 48.71 in the the number of nearest neighbor points. The larger the number of
ITR. Nevertheless, there is no significant difference between the nearest neighbor points, the higher the recognition accuracy, but the
CAR-USSR, PCA-USSR, MDS-USSR, and LLE-USSR methods (p > 0.05). processing time also increases exponentially. The above results suggest
As a representative of nonlinear manifold learning methods, the that the CCA-USSR method is suitable for real-time SSVEP detection
recognition accuracy of the LLE-USSR method greatly depends on the technology and neuroscience.
number of nearest neighbors. Meanwhile, from the experimental
results, it can be seen that although the number of nearest neighbors
is set to 40, the LLE-USSR method has not achieved competitive BCI 4. Discussion
performance. Compared with other dimensionality reduction
methods, CCA can retain the most features of multi-channel SSVEP This study discusses the impact of different dimensionality
signals and the CCA-USSR method has the highest recognition reduction methods on multi-channel SSVEP signals. Five typical
accuracy, ITR, and robustness (p < 0.05). Meanwhile, compared with unsupervised methods were compared from the perspective of the
typical CCA and FBCCA methods, the CCA-USSR method is also CCA coefficient spectrum, recognition accuracy, ITR, robustness, and
more suitable for SSVEP signals induced by the motion checkerboard processing speed. In the CAR method, the selection of channels is
paradigm (p < 0.05). Hence, we can conclude that CCA is currently the particularly important. The experimental results found that among
best dimensionality reduction method in EEG signals, and the these eight electrodes POz, PO3, PO4, PO5, PO6, Oz, O1, and O2,
untrained CCA-USSR method can achieve satisfactory results in real- using Oz as the reference channel can achieve the highest recognition
time BCI applications and spectral analysis. accuracy and ITR. This also indicates that the Oz channel which is

Frontiers in Neuroscience 07 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

TABLE 1 Detection performance using the five methods (T = 2 s).

Subjects Accuracy % ITR bits/ Accuracy % ITR bits/ Accuracy % ITR bits/ Accuracy % ITR bits/
(CCA) min (CCA) (FBCCA) min (CAR- min (CAR- (PCA- min (PCA-
(FBCCA) USSR) USSR) USSR) USSR)
S1 94.29 135.68 82.86 107.89 8.57 1.68 54.29 54.27

S2 28.57 18.97 34.29 25.76 80.00 101.70 42.86 37.11

S3 77.14 95.73 68.57 78.97 74.29 89.96 74.29 89.96

S4 48.57 45.40 54.29 54.27 57.14 58.91 48.57 45.40

S5 88.57 121.05 65.71 73.72 54.29 54.27 22.86 12.87

S6 94.29 135.68 77.14 95.73 88.57 121.05 77.14 95.73

S7 74.29 89.96 82.86 107.89 20.00 10.12 57.14 58.91

S8 77.14 95.73 65.71 73.72 5.71 0.50 62.86 68.64

S9 68.57 78.97 60.00 63.70 88.57 121.05 54.29 54.27

S10 34.29 25.76 40.00 33.18 77.14 95.73 37.14 29.39

S11 42.86 37.11 57.14 58.91 62.86 68.64 45.71 41.18

S12 34.29 25.76 68.57 78.97 25.71 15.83 71.43 84.38

S13 31.43 22.28 45.71 41.18 28.57 18.97 65.71 73.72

S14 57.14 58.91 65.71 73.72 77.14 95.73 60.00 63.70

S15 42.86 37.11 71.43 84.38 28.57 18.97 71.43 84.38

S16 62.86 68.64 62.86 68.64 82.86 107.89 85.71 114.32

S17 74.29 89.96 62.86 68.64 82.86 107.89 77.14 95.73

S18 62.86 68.64 65.71 73.72 74.29 89.96 57.14 58.91

S19 85.71 114.32 74.29 89.96 94.29 135.68 65.71 73.72

S20 74.29 89.96 65.71 73.72 2.86 0 37.14 29.39

S21 60.00 63.70 65.71 73.72 51.43 49.76 77.14 95.73

S22 68.57 78.97 68.57 78.97 94.29 135.68 65.71 73.72

S23 65.71 73.72 71.43 84.38 88.57 121.05 48.57 45.40

S24 82.86 107.89 91.43 128.14 85.71 114.32 74.29 89.96

S25 51.43 49.76 60.00 63.70 45.71 41.18 37.14 29.39

S26 74.29 89.96 74.29 89.96 31.43 22.28 65.71 73.72

S27 62.86 68.64 60.00 63.70 8.57 1.68 60.00 63.70

S28 68.57 78.97 71.43 84.38 25.71 15.83 42.86 37.11

S29 40.00 33.18 80.00 101.70 82.86 107.89 57.14 58.91

S30 11.43 3.32 40.00 33.18 8.57 1.68 45.71 41.18

S31 62.86 68.64 68.57 78.97 2.86 0 37.14 29.39

S32 91.43 128.14 77.14 95.73 94.29 135.68 82.86 107.89

S33 40.00 33.18 77.14 95.73 85.71 114.32 28.57 18.97

S34 85.71 114.32 77.14 95.73 17.14 7.59 20.00 10.12

S35 71.43 84.38 77.14 95.73 57.14 58.91 60.00 63.70

S36 68.57 78.97 68.57 78.97 2.86 0 62.86 68.64

S37 54.29 54.27 65.71 73.72 82.86 107.89 54.29 54.27

S38 51.43 49.76 60.00 63.70 71.43 84.38 68.57 78.97

S39 40.00 33.18 62.86 68.64 8.57 1.68 31.43 22.28

S40 42.86 37.11 54.29 54.27 77.14 95.73 17.14 7.59

S41 57.14 58.91 65.71 73.72 5.71 0.50 40.00 33.18

S42 62.86 68.64 62.86 68.64 57.14 58.91 62.86 68.64

Mean ± SD 61.16 ± 19.59 66.68 ± 33.48 65.99 ± 11.68 74.22 ± 20.66 52.38 ± 32.75 51.25 ± 48.71 54.97 ± 17.38 55.36 ± 27.32

(Continued)

Frontiers in Neuroscience 08 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

TABLE 1 (Continued)

Subjects Accuracy % ITR bits/min Accuracy % ITR bits/min Accuracy % ITR bits/min
(MDS-USSR) (MDS-USSR) (LLE-USSR) (LLE-USSR) (CCA-USSR) (CCA-USSR)
S1 57.14 58.91 65.71 73.72 85.71 114.32

S2 34.29 25.76 25.71 15.83 54.29 54.27

S3 74.29 89.96 54.29 54.27 74.29 89.96

S4 54.29 54.27 62.86 68.64 57.14 58.91

S5 57.14 58.91 40.00 33.18 77.14 95.73

S6 82.86 107.89 65.71 73.72 88.57 121.05

S7 60.00 63.70 42.86 37.11 82.86 107.89

S8 60.00 63.70 51.43 49.76 71.43 84.38

S9 54.29 54.27 62.86 68.64 62.86 68.64

S10 37.14 29.39 42.86 37.11 51.43 49.76

S11 51.43 49.76 28.57 18.97 60.00 63.70

S12 71.43 84.38 60.00 63.70 77.14 95.73

S13 65.71 73.72 45.71 41.18 65.71 73.72

S14 54.29 54.27 40.00 33.18 77.14 95.73

S15 77.14 95.73 45.71 41.18 80.00 101.70

S16 85.71 114.32 77.14 95.73 88.57 121.05

S17 82.86 107.89 51.43 49.76 77.14 95.73

S18 57.14 58.91 37.14 29.39 62.86 68.64

S19 62.86 68.64 42.86 37.11 88.57 121.05

S20 68.57 78.97 42.86 37.11 74.29 89.96

S21 68.57 78.97 68.57 78.97 77.14 95.73

S22 65.71 73.72 62.86 68.64 82.86 107.89

S23 45.71 41.18 51.43 49.76 65.71 73.72

S24 77.14 95.73 65.71 73.72 88.57 121.05

S25 40.00 33.18 51.43 49.76 65.71 73.72

S26 71.43 84.38 60.00 63.70 80.00 101.70

S27 65.71 73.72 45.71 41.18 80.00 101.70

S28 42.86 37.11 42.86 37.11 80.00 101.70

S29 62.86 68.64 68.57 78.97 82.86 107.89

S30 51.43 49.76 28.57 18.97 54.29 54.27

S31 45.71 41.18 25.71 15.83 77.14 95.73

S32 80.00 101.70 65.71 73.72 88.57 121.05

S33 28.57 18.97 37.14 29.39 68.57 78.97

S34 28.57 18.97 17.14 7.59 82.86 107.89

S35 62.86 68.64 28.57 18.97 85.71 114.32

S36 71.43 84.38 57.14 58.91 74.29 89.96

S37 57.14 58.91 54.29 54.27 80.00 101.70

S38 74.29 89.96 62.86 68.64 71.43 84.38

S39 25.71 15.83 28.57 18.97 65.71 73.72

S40 22.86 12.87 57.14 58.91 68.57 78.97

S41 54.29 54.27 25.71 15.83 60.00 63.70

S42 68.57 78.97 42.86 37.11 71.43 84.38

Mean ± SD 58.57 ± 16.29 61.29 ± 26.45 48.44 ± 14.73 45.20 ± 21.77 74.01 ± 10.42 89.42 ± 20.35

Frontiers in Neuroscience 09 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

FIGURE 5
(A) Average accuracy using the CCA recognition method. (B) Average ITR using the CCA recognition method. (C) Average accuracy using the PSDA
recognition method. (D) Average ITR using the PSDA recognition method.

TABLE 2 Comparison of the processing time of five unsupervised the features of SSVEP data. In terms of processing time, CAR, PCA, and
dimensionality reduction methods.
CCA are promising for real-time BCI detection technology, while MDS
Data length and LLE need more time for processing BCI.
Only when the dimensionality reduction method retains the most
3s 2.5 s 2s 1.5 s 1s
features of the original multi-channel signal, the USSR model can
CAR 0.0013 s 0.0012 s 0.0012 s 0.0011 s 0.0010 s more effectively utilize synergistic effects to enhance the energy and
PCA 0.0025 s 0.0023 s 0.0021 s 0.0019 s 0.0017 s amplitude of target frequencies, thereby increasing ITR. The reason
MDS 5.1042 s 2.7754 s 1.4109 s 0.4317 s 0.1669 s why SR is different from other traditional denoising methods is that
the SR model utilizes a dynamic feature enhancement mechanism
LLE 9.9095 s 6.2133 s 3.6962 s 1.1774 s 0.3849 s
with the help of the synergetic action of input aperiodic signal, noise,
CCA 0.0637 s 0.0615 s 0.05942 s 0.05776 s 0.0543 s and the nonlinear resonance system. The SR model considers noise as
a positive factor, thereby using noise energy to enhance weak signal
located over the occipital region retains the most effective features of features. In addition, compared with the first-order bistable stochastic
SSVEP signals. In linear dimensionality reduction methods such as resonance, the underdamped second-order SR considers the inertia
CAR, PCA, MDS, and CCA, although they have different rules for term and normalizes the damping factor. To use an analogy, first-order
projecting high-dimensional data to low-dimensional space, CCA BSR processing means primary filtering and second-order USSR
shows the best adaptability to SSVEP signals and MDS comes next. In processing means secondary filtering, thereby producing a cleaner
nonlinear LLE methods, the number of nearest neighbor points filtered response than first-order SR. This is the reason why the USSR
determines the quality of detection results and processing speed. The can more effectively improve the weak signal detection performance
larger the number of nearest neighbor points, the higher the recognition than BSR and other classic linear methods.
accuracy and ITR, and the longer the processing time. However, due to Although the motion checkerboard paradigm can effectively
the strong limitations of the LLE method, which assumes that the input reduce the fatigue of subjects and is more suitable for long-time BCI
data satisfies the manifold distribution, it is not suitable for extracting performance detection, its evoked SSVEP signal has fewer harmonic

Frontiers in Neuroscience 10 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

components and more concentrated frequency energy (Han et al., Ethics statement
2018). Therefore, some standard untrained algorithms, such as CCA
and FBCCA, are not effective in detecting the subject’s purpose (Yan Ethical approval was not required for the study involving human
et al., 2021). In addition, using some traditional linear methods to samples in accordance with the local legislation and institutional
decode SSVEP signals, the useful features will be attenuated or lost requirements because [reason ethics approval was not required].
while denoising, which seriously affects the improvement of Written informed consent for participation in this study was provided
recognition accuracy. Nevertheless, the nonlinear SR model has a by the participants’ legal guardians/next of kin.
frequency energy concentration effect similar to the motion
checkerboard pattern. And among typical SR models, the USSR model
has the best performance. Naturally, the study proposes to combine Author contributions
dimensionality reduction methods and nonlinear USSR models to
extract non-stationary SSVEP features. The experimental results RC: conceptualization, methodology, software, formal analysis,
indicate that among traditional dimensionality reduction methods, data curation, writing—original draft, and writing—review and
the CCA method still has the projecting rule that best matches the editing. GX: validation, formal analysis, and investigation. HZ:
SSVEP rhythm, which can retain the most original multi-channel software. XZ and BL: supervision. JW: funding acquisition. SZ: project
SSVEP features, and the USSR model can effectively highlight the administration. All authors contributed to the article and approved
energy and amplitude of the target frequency with noise immunity, the submitted version.
thereby increasing the algorithm robustness, recognition accuracy,
and ITR. This untrained method also provides the possibility of
applying a nonlinear model from one-dimensional signals to multi- Funding
dimensional signals.
This work was supported in part by the key Research and
Development Projects of Shaanxi Province under grant
5. Conclusion no.2021ZD0204300, in part by the Science and Technology Plan
Project of Xi’an under grant 20KYPT0001-10, and in part by the Key
In this study, we first compare five typical unsupervised Research and Development Program of Shaanxi Province of China
dimensionality reduction methods, namely CAR, PCA, MDS, LLE, and under grant 2021GXLH-Z-008.
CCA. The experimental results show that CCA has the highest
adaptability for SSVEP rhythms and can retain the most effective
features. Furthermore, compared with the standard CCA and FBCCA Conflict of interest
methods, the novel untrained CCA-USSR method proposed in this
paper can more effectively highlight the target frequency and have The authors declare that the research was conducted in the
higher robustness, thereby increasing recognition accuracy and ITR. In absence of any commercial or financial relationships that could
addition, the CCA-USSR method also has advantages in processing be construed as a potential conflict of interest.
speed and has the potential for real-time BCI detection technology.

Publisher’s note
Data availability statement
All claims expressed in this article are solely those of the authors
The data analyzed in this study is subject to the following licenses/ and do not necessarily represent those of their affiliated organizations,
restrictions: the data is available from the corresponding author upon or those of the publisher, the editors and the reviewers. Any product
reasonable request. Requests to access these datasets should that may be evaluated in this article, or claim that may be made by its
be directed to GX, ghxu@mail.xjtu.edu.cn. manufacturer, is not guaranteed or endorsed by the publisher.

References
Benzi, R., Sutera, A., and Vulpiani, A. (1981). The mechanism of stochastic resonance. detection technology. Expert Syst. Appl. 225:120141. doi: 10.1016/j.eswa.2023.
J. Phys. A Math. Gen. 14, L453–L457. doi: 10.1088/0305-4470/14/11/006 120141
Chen, T. P. J., and Gao, X. (2015). Filter bank canonical correlation analysis for Chen, R., Xu, G., Zhang, X., Han, C., and Zhang, S. (2021a). Multi-scale noise transfer
implementing a high-speed SSVEP-based brain-computer interface. J. Neural Eng. and feature frequency detection in SSVEP based on FitzHugh–Nagumo neuron system.
12:046008. doi: 10.1088/1741-2560/12/4/046008 J. Neural Eng. 18:056054. doi: 10.1088/1741-2552/ac2bb7
Chen, X., Wang, Y., Nakanishi, M., Gao, X., Jung, T. P., and Gao, S. (2015). High-speed Chen, R., Xu, G., Zheng, Y., Yao, P., Zhang, S., Yan, L., et al. (2021b). Waveform feature
spelling with a noninvasive brain–computer interface. Proc. Natl. Acad. Sci. 112, E6058– extraction and signal recovery in single-channel TVEP based on Fitzhugh–Nagumo
E6067. doi: 10.1073/pnas.1508080112 stochastic resonance. J. Neural Eng. 18:056031. doi: 10.1088/1741-2552/ac2459
Chen, R., Xu, G., Jia, Y., Zhou, C., Wang, Z., Pei, J., et al. (2022). Enhancement of Collins, J., Chow, C. C., and Imhoff, T. T. (1995). Aperiodic stochastic resonance in
time-frequency energy for the classification of motor imagery electroencephalogram excitable systems. Phys. Rev. E 52, R3321–R3324. doi: 10.1103/PhysRevE.52.R3321
based on an improved FitzHugh–Nagumo neuron system. IEEE Trans. Neural Syst.
De Ridder, D., Kouropteva, O., Okun, O., and Duin, R. (2003). Supervised locally linear
Rehabil. Eng. 31, 282–293. doi: 10.1109/TNSRE.2022.3219450
embedding. In: Artificial neural networks and neural information processing—ICANN/
Chen, R., Xu, G., Pei, J., Gao, Y., Zhang, S., and Han, C. (2023). Typical stochastic ICONIP 2003: Joint international conference ICANN/ICONIP 2003 Istanbul, Turkey, June
resonance models and their applications in steady-state visual evoked potential 26–29, 2003 proceedings. Springer.

Frontiers in Neuroscience 11 frontiersin.org

Chen et al. 10.3389/fnins.2023.1246940

Fauve, S., and Heslot, F. (1983). Stochastic resonance in a bistable system. Phys. Lett. Ming, C., Xiaorong, G., Shangkai, G., and Dingfeng, X. (2002). Design and
A 97, 5–7. doi: 10.1016/0375-9601(83)90086-5 implementation of a brain-computer interface with high transfer rates. I.E.E.E. Trans.
Biomed. Eng. 49, 1181–1186. doi: 10.1109/TBME.2002.803536
Friman, O., Volosyak, I., and Graser, A. (2007). Multiple channel detection of steady-
state visual evoked potentials for brain-computer interfaces. IEEE transactions on Nakanishi, M., Wang, Y., Chen, X., Wang, Y. T., Gao, X., and Jung, T. P. (2017).
biomedical engineering. 54, 742–750. doi: 10.1109/TBME.2006.889160 Enhancing detection of SSVEPs for a high-speed brain speller using task-related
component analysis. I.E.E.E. Trans. Biomed. Eng. 65, 104–112. doi: 10.1109/
Han, C., Xu, G., Xie, J., Chen, C., and Zhang, S. (2018). Highly interactive brain– TBME.2017.2694818
computer Interface based on flicker-free steady-state motion visual evoked potential.
Sci. Rep. 8:5835. doi: 10.1038/s41598-018-24008-8 Nakanishi, M., Yijun, W., and Tzyy-Ping, J. (2018). Spatial filtering techniques for
improving individual template-based SSVEP detection. Signal Process Machine Learn
Hu, S., Xu, C., Guan, W., Tang, Y., and Liu, Y. (2014). Texture feature extraction based Brain-Machine Interfaces, 219–242. doi: 10.1049/PBCE114E_ch11
on wavelet transform and gray-level co-occurrence matrices applied to osteosarcoma
diagnosis. Biomed Mater Eng. 24, 129–143. doi: 10.3233/BME-130793 Orekhova, E. V., Wallin, B. G., and Hedstr M, A. (2002). Modification of the average
reference montage: dynamic average reference. J. Clin. Neurophysiol. 19, 209–218. doi:
Jin, J., Wang, Z., Xu, R., Liu, C., Wang, X., and Cichocki, A. (2021). Robust similarity 10.1097/00004691-200206000-00004
measurement based on a novel time filter for SSVEPs detection. IEEE Trans Neural Netw
Learn Syst 34, 4096–4105. doi: 10.1109/TNNLS.2021.3118468 Saeed, N., Nam, H., Haq, M. I. U., and Muhammad Saqib, D. B. (2018). A survey on
multidimensional scaling. ACM Computing Surveys (CSUR). 51, 1–25. doi: 10.1145/3178155
Kramer, D. R., Lee, M. B., Barbaro, M. F., Gogia, A. S., Peng, T., Liu, C. Y., et al. (2021).
Mapping of primary somatosensory cortex of the hand area using a high-density Vidal, J. J. (1973). Toward direct brain-computer communication. Annu. Rev. Biophys.
electrocorticography grid for closed-loop brain computer interface. J. Neural Eng. Bioeng. 2, 157–180. doi: 10.1146/annurev.bb.02.060173.001105
18:036009. doi: 10.1088/1741-2552/ab7c8e Wold, S., Esbensen, K., and Geladi, P. (1987). Principal component analysis. Chemom.
Li, R., Liu, D., Li, Z., Liu, J., Zhou, J., Liu, W., et al. (2022). A novel EEG decoding Intel. Lab. Syst. 2, 37–52. doi: 10.1016/0169-7439(87)80084-9
method for a facial-expression-based BCI system using the combined convolutional Wu, D., and Zhu, S. (2008). Stochastic resonance in FitzHugh–Nagumo system with
neural network and genetic algorithm. Front. Neurosci. 16:988535. doi: 10.3389/ time-delayed feedback. Phys. Lett. A 372, 5299–5304. doi: 10.1016/j.physleta.2008.06.015
fnins.2022.988535
Yan, W., du, C., Luo, D., Wu, Y. C., Duan, N., Zheng, X., et al. (2021). Enhancing
Li, Y., Wang, F., Chen, Y., Cichocki, A., and Sejnowski, T. (2017). The effects of detection of steady-state visual evoked potentials using channel ensemble method. J.
audiovisual inputs on solving the cocktail party problem in the human brain: an fMRI Neural Eng. 18:046008. doi: 10.1088/1741-2552/abe7cf
study. Cereb. Cortex 28, 3623–3637. doi: 10.1093/cercor/bhx235
Yao, P., Xu, G., Jia, L., Duan, J., Han, C., Tao, T., et al. (2019). Multiscale noise
Lin, Z., Zhang, C., Wu, W., and Gao, X. (2006). Frequency recognition based on suppression and feature frequency extraction in SSVEP based on underdamped second-
canonical correlation analysis for SSVEP-based BCIs. I.E.E.E. Trans. Biomed. Eng. 53, order stochastic resonance. J. Neural Eng. 16:036032. doi: 10.1088/1741-2552/ab16f9
2610–2614. doi: 10.1109/TBME.2006.886577
Zhang, Y., Dong, L., Zhang, R., Yao, D., Zhang, Y., and Xu, P. (2014c). An efficient
Liu, B., Chen, X., Shi, N., Wang, Y., Gao, S., and Gao, X. (2021). Improving the frequency recognition method based on likelihood ratio test for SSVEP-based BCI.
performance of individually calibrated SSVEP-BCI by task-discriminant component Comput. Math. Methods Med. 2014:908719. doi: 10.1155/2014/908719
analysis. IEEE Trans. Neural Syst. Rehabil. Eng. 29, 1998–2007. doi: 10.1109/
TNSRE.2021.3114340 Zhang, Y., Xu, P., Cheng, K., and Yao, D. (2014b). Multivariate synchronization index
for frequency recognition of SSVEP-based brain–computer interface. J. Neurosci.
Lu, S., He, Q., and Kong, F. (2015). Effects of underdamped step-varying second-order Methods 221, 32–40. doi: 10.1016/j.jneumeth.2013.07.018
stochastic resonance for weak signal detection. Digit. Signal Process. 36, 93–103. doi:
10.1016/j.dsp.2014.09.014 Zhang, Y., Zhou, G., Jin, J., Wang, X., and Cichocki, A. (2014a). Frequency recognition
in SSVEP-based BCI using multiset canonical correlation analysis. Int. J. Neural Syst.
Luce, R. D., and Perry, A. D. (1949). A method of matrix analysis of group structure. 24:1450013. doi: 10.1142/S0129065714500130
Psychometrika 14, 95–116. doi: 10.1007/BF02289146
Zhang, Y., Zhou, G., Zhao, Q., Onishi, A., Jin, J., Wang, X., et al. (2011). Multiway
Mcnamara, B., Wiesenfeld, K., and Roy, R. (1988). Observation of stochastic resonance canonical correlation analysis for frequency components recognition in SSVEP-based
in a ring laser. Phys. Rev. Lett. 60, 2626–2629. doi: 10.1103/PhysRevLett.60.2626 BCIs. In: International conference on neural information processing. Springer.

Frontiers in Neuroscience 12 frontiersin.org