Twenty-seven healthy subjects (19 males and 6 females, two subjects were excluded because of poor performance of fixing their eyes or head movements during the attention period; ages range: 21–38 years old) were recruited to participate in the experiment. All of them provided informed consent for their data to be published. This study was approved by the Ethics Committee of Sichuan Provincial Rehabilitation Hospital (approval number: CKLL-2018008), which is our cooperating institution. All participants had normal or corrected-to-normal vision.

EEG data were recorded from 64 channels using electrodes placed according to the standard 10-20 electrode system using a Synamps2 system (Neuroscan, Inc.) with a sampling rate of 250 Hz. The electrodes placed on the forehead (GND) and right mastoid (A2) were used as the ground and reference electrodes, respectively. The 24 channels in the occipital region, including CP1, CPz, CP2, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO5, PO3, POz, PO4, PO6, PO8, CB1, O1, Oz, O2, and CB2 based on the standard positions in the 10-20 system, were chosen for further analysis since we were only concerned with vision-related activity (Chen et al., 2015b). The impedances of the electrodes were maintained below 10 kΩ. Eye movements were controlled for with two additional channels that recorded the horizontal electrooculogram. The experiments were presented on an LCD monitor at 60 Hz and a screen resolution of 1920 × 1080. Participants were seated in a dimly lit room and kept their eyes approximately 60 cm from the screen.

First, 70 trials were presented during the covert attention phase. As depicted in Figure 1A, at the beginning of each trial, the participants were instructed to fixate on a cross at the center of the screen. Then, a numerical cue appeared for 2 s on the top of the cross, indicating the upcoming 8-s period of covert attention, in which the participants had to shift their attention to a stimulus target that flickered at a frequency of 7.5 Hz while strictly maintaining focus on the center cross. It should be noted that familiar with the instruction before the experiments especially the covert attention experiment was necessary for all participants. In our approach, in order to maintain the focusing state and during stimulation, 2 s of preparation was also provided for participants to move their gaze to the central cross on the screen before the onset of peripheral stimulation on each trial. Besides, the severe eye movements were detected and the corresponding trials were rejected by manual proofreading of the recorded EOG signals. Moreover, we also monitored the head movement in the thole procedure of each trial simultaneously. The participants need to provide feedback of their concentrate performance when the experiments were completed. The shape of the stimulus target was a circle of size 0.75° (i.e., a radius of 7.81 mm) whose position was randomly distributed in a circular area of size 13.90° (i.e., a radius of 148 mm).

All subjects were also required to participate in a controlling experiment (i.e., overt attention) involving 20 trials for each subject. The only difference with the covert attention experiment trials was that the fixation cross was removed; instead, each participant was asked to fix his or her attention on the flickering stimulus circle during each overt attention trial, in which the duration of stimulus presentation was kept the same (i.e., 8 s). Subsequently, an additional 20 trials without attentional requirements were collected for each participant for the subsequent classification analysis. Specifically, people need to remain idle state (i.e., they were allowed to freely blink as usual, and perform some basis physiological activities but avoid vigorous body movements) during the experiments of no attention.

A 2-s interval was presented between continuous trials for both the covert and overt attention experiments, during which no flickering circles were displayed on the screen. The onset of each trial was cued by a transitory text prompt. During the attention period of each trial, the subjects were instructed to refrain from blinking as much as possible. To keep the subjects engaged, they were required to count every second until the end of each stimulus period.

To assess the SSVEP response in the covert attention task for different degrees of spatial distance, we divided the eccentricities into three levels as follows. Eccentricities not greater than 3° (i.e., Euclidean distance ≤ 31.45mm in this study) was defined as “near,” while eccentricities not less than 10° (i.e., Euclidean distance ≥105.8mm in this study) was defined as “moderate,” and the remaining eccentricities were defined as “far” (Figure 1B).

Raw EEG signals were acquired from the 64 channels using electrodes placed according to the standard 10-20 electrode system using a SynAmpsRT amplifier (Neuroscan Compumedics, Singen, Germany) at a sampling frequency of 250 Hz. The EEG signals were bandpass filtered between 4 and 25 Hz using a 4th − order Butterworth filter prior to further analysis. Trials containing signal artifacts corresponding to behaviors such as jaw clenching, muscle movement, and electrode disconnections were removed by visual inspection. Next, independent component analysis (ICA) was performed to remove blinks and eye movements using the FieldTrip toolbox (Version 20201021, http://www.fieldtriptoolbox.org) with default parameters. Components with upper 70% confidence labeled as eye artifacts were then removed. All analyses were performed in Matlab.

This study employed the extended canonical correlation analysis (CCA)-based method, using the multiple trials of overt attention as individual calibration signals to extract the metrical features for the test signals during covert attention (Nakanishi et al., 2014, 2015; Chen et al., 2015b). Here, with Y representing the sinusoidal template data (Chen et al., 2015b), the individual calibration signals for the nth subband component were denoted by a three-way tensor χ^∈ℝNn×Nc×Ns, and the test signals of each trial during covert attention were denoted by a four-way tensor X∈ℝNt×Nn×Nc×Ns. Note that N_t indicates the number of stimulation trials during the covert attention experiment, N_n indicates the number of subband components, N_c is the number of channels and N_s is the number of sampling points in each trial. Filter bank analysis was applied to decompose the SSVEPs into subband components so that the harmonic components could be extracted more efficiently (Chen et al., 2015a). In practice, Butterworth infinite impulse response (IIR) bandpass filters that shared the same upper-bound frequency (30 Hz) but had different lower-bound frequencies (i.e., for the nth subband component X_n, the lower-bound frequency is 4+(n−1)×6 Hz) were used to extract the harmonic components (i.e., N_n = 3 in this study) from the test EEG data X. Next, the following weight vectors were used as spatial filters to enhance the signal-to-noise ratio (SNR) of the SSVEPs: (i) U_Xn and V_X were calculated using the nth subband component of the test EEG data epoch X_n and the sine-cosine template data epoch Y. (ii) UXnX^n and VXnX^n were calculated using X_n and the nth subband component of the individual calibration data, X^n. (iii) UX^nY and VX^nY were calculated from X^n,k and Y.

Then, a vector of correlation coefficients was obtained using these three pairs of weight vectors (Chen et al., 2015b; Nakanishi et al., 2015), as shown below:

where ρ(a, b) denotes the correlation coefficient between a and b. Note that both Y and X^n are previously determined data epochs associated with the kth target. Thus, the feature of the nth subband of the test data epoch X_n associated with the kth target was defined in terms of r_n:

where sign() is used to preserve the discriminating information because negative correlation coefficients might exist between the two signals being compared. Furthermore, a weighting coefficient w(n) was applied to each ρ_n,k because the SNR of the harmonic components of the SSVEP decreases as the evoked frequency increases. The weight coefficients are defined as follows:

where a and b are constants determined using a grid-search method during offline analysis (1 and 0, respectively, in this study). Finally, the synthetic correlation parameter ρ to specific eccentricity of each attentional trial associated with test data X, is defined as a weighted square sum of the correlative features ρ_n of all subbands (Chen et al., 2015b; Nakanishi et al., 2015) as follows:

The PSD of the EEG data was calculated by performing fast Fourier transform. The fundamental and two harmonic components of the stimulative frequency were selected as the frequencies of interest (i.e., 7.5 Hz, 15 Hz and 22.5 Hz) during the calculation. Moreover, the ITPC was calculated to visualize the trial-by-trial variations in the time-frequency domain (Yu et al., 2018; Gui et al., 2020). Specifically, the single-trial EEG data were transformed into the frequency domain from 4 to 25 Hz in steps of 0.5 Hz using wavelet transform with a sliding time window from 0 to 8 s in steps of 0.05 s. The coefficient of the wavelet transform was denoted as X_k(f) for the k^th trial (k = 1, 2, …, n), and the ITPC is defined below:

where A_k(f) = ∠X_k(f) is the phase information obtained from wavelet analysis.

In addition to the PSD, network patterns can be used for extracting important physiological information on the way different brain regions are functionally coupled (La Rocca et al., 2014). In this study, the network construction was estimated by calculating spectral coherence (COH), which is frequently used due to its practical and intuitive interpretation. Given two EEG signals from channels i and j, the spectral coherence Coh_i,j(f) for a particular frequency f is defined as

where P_i,j(f) denotes the cross-spectra of the signals corresponding to channels i and j, while P_i,i(f) and P_j,j(f) are the respective autospectra. To construct the network, we used the 62 electrodes as network nodes and applied COH to calculate the connectivity weights for all overt and covert attention trials.

The overt, covert, and no attention trials were used for classification analysis separately. First, we used the above correlation values ρ calculated by FBCCA corresponding to each kind of attention trail as features for three-class classification. The k-nearest neighbors (K-NN) algorithm (Hosseini et al., 2020) was chosen as the classifier. Please refer to Figure 2 for an overview of the procedure for classification. In addition, the same classifier was used to identify which type of covert attention the subjects were attending (i.e., near, moderate, and far), with the correlation values corresponding to all three types of covert trials being selected as features. We conducted 10-fold cross-validation (i.e., 90% of the data were used as a training set and the remaining 10% of the data as a test set) to assess the classification performance for each classification and each subject. The statistical significance of the classification accuracy was evaluated through the Wilcoxon signed-rank test.

The correlation values and p values of the statistical tests for each subject are given in Figure 3. Furthermore, to investigate the subjects' SSVEP response with eccentricity under different visual angles, we also compared the relationship between the synthetic correlation parameter ρ and both the horizontal distance (Figure 4) and vertical distance (Figure 5). In general, the subjects generated similar characteristics of SSVEPs (i.e., across subjects, the parameter in the covert attention condition was significantly correlated with visual target eccentricity).

For the majority of subjects (except for S8, S16, and S19, Figure 3), there were negative and significant associations between the parameter in the covert attention task and the eccentricity as represented using the Euclidean distance. In addition, only six (S4, S13, S16, S21, S22, and S25, Figure 4) and eleven (S3, S6, S8, S9, S10, S13, S16, S17, S18, S19, and S12, Figure 5) of all subjects did not show significant correlations with the horizontal distance or the vertical distance were applied, respectively. Therefore, for each subject, a statistically significant negative association between the feature for covert attention and eccentricity can be found if the proper eccentricity measurement was selected (i.e., calculating the eccentricity either with the Euclidean, horizontal or vertical distance). Moreover, one-way repeated measures ANOVA revealed a significant trend of group effect [F_(2,72) = 15.4, p < 0.001] on the three different types of eccentricity.

The EEG topography result for the average of all subjects under the overt condition and different levels of covert attention can be compared in Figure 6. When subjects directed their attention covertly to the visual target rather than to the fixation cross, the SSVEP amplitudes at the stimulus frequency decreased over the area near the occipital and parietal lobes [F_(1,48) = 22.43, p < 0.0001]. Additionally, covert attention with the flickering target further from the center cross had a narrower power distribution than that with targets closer to the cross. A significant main effect of the three groups [F_(2,72) = 3.62, p = 0.032] over the three different types of covert attention was found.

We also compared the difference in EEG power across time-frequency windows between these two experiments, as shown in Figure 7A. The data were standardized per channel of each trial, and outliers were removed before the frequency and time-frequency analysis. In the spectral domain, it can be seen that the amplitude of the fundamental frequency, as well as the second-harmonic frequency and the third-harmonic frequency in overt attention, was greater than each of that in different covert attention [F_(1,48) = 56.59, p < 0.0001]. For the details of covert attention, the amplitude at the three frequencies of interest showed a decreasing trend from the near condition to the far condition [F_(2,72) = 2.71, p = 0.073 at 7.5 Hz, F_(2,72) = 13.38, p < 0.001 at 15.0 Hz, F_(2,72) = 20.65, p < 0.001 at 22.5 Hz].

In addition, compared with that in the overt condition, the ITPC in the covert condition (Figure 7B) was reduced [F_(1,48) = 20.77 ~ 74.01, p < 0.0001 at 7.5 Hz, F_(1,48) = 41.88 ~ 143.81, p < 0.0001 at 15.0 Hz]. Additionally, a significant main effect of the three groups over the three different types of covert attention across the whole was found [F_(2,72) = 20.77 ~ 45.80, p < 0.0001 at 7.5 Hz, F_(2,72) = 30.28 ~ 88.50, p < 0.0001 at 15.0 Hz].

Figure 8 depicts the linkages with significant differences (p < 0.05) among the overt and covert attention conditions for all subjects as revealed by paired-sample t-test. The results show that significantly different connections mainly occurred between the nodes located in the occipital lobes. However, no significant differences were found in the linkages between the different groups of covert attention, i.e., between near and moderate, between near and far, and between moderate and far. This conclusion is consistent with an fMRI study (Beauchamp et al., 2001) that revealed that more activity was observed in the precentral sulcus during overt attention than during covert attention. Another finding of this study is similar to that of the aforementioned fMRI study (Liu et al., 2003): when the subjects tended to pay attention to peripheral stimuli, the active network did not extend into the superior frontal or inferior frontal areas of the brain.

Figure 9 displays the confusion matrices averaged across subjects. The diagonal cells show the number of cases that were correctly classified, and the off-diagonal cells show the misclassified cases. In total, the average accuracy of 91.42±5.30% for the three classes between the overt, covert and no attention (Figure 9A) was obtained across all subjects, as entries on the main diagonal indicate that the most prominent features were identified. Large values were also observed on the superdiagonal and subdiagonal lines between covert attention and unattended attention, suggesting that these two conditions are relatively easily misclassified as each other. As for the decode performance of different types of covert attention, we acquired a average accuracy of 66.67±14.21% across all subjects (Figure 9B). Values on the main diagonal demonstrated that the most prominent features for the types of near and moderate condition were recognized except the type of far condition as most trails were misclassified into the moderate condition.