Neurofeedback training of control network improves SSVEP-based BCI performance in children

Background: In the past 20 years, neural engineering has made unprecedented progress in the interpretation of brain information (e.g., brain-computer interfaces) and neuromodulation (e.g., electromagnetic stimulation and neurofeedback). However, the study of improving the performance of the brain-computer interface (BCI) using the neuromodulation method rarely exists. The present study designs a neurofeedback training method to improve the performance of steady-state visual evoked potential (SSVEP) BCI and further explores its underlying mechanisms. Methods: As parietal lobe is the sole hub of information transmission, up-regulating alpha-band power of the parietal lobe by neurofeedback training was present as a new neural modulating method to improve SSVEP-based BCI in this study. Results: After this neurofeedback training (NFT), the signal-to-noise ratio (SNR), accuracy, and information transfer rate (ITR) of SSVEP-based BCI were increased by 5.8%, 4.7%, and 15.6% respectively. However, no improvement has been observed in the control group in which the subjects do not participate in NFT. What’s more, a general reinforcement of the information flow from the parietal lobe to the occipital lobe was also observed. Evidence from the network analysis and attention test further indicate that NFT improves attention via developing the control ability of the parietal lobe and then enhances the above SSVEP indicators. Conclusion: Up-regulating parietal alpha-amplitude using neurofeedback training significantly improves the SSVEP-baesd BCI performance through modulating the control network. The study validates an effective neuromodulation method, and possibly also contributes to explaining the function of the parietal lobe in the control network.


Introduction 3
The study of cortical oscillations observed in the electroencephalogram (EEG) that have made significant progress recently. The effects and influences of various rhythms on people have gradually been researched since the neural oscillations were first discovered by Hans Berger (Berger, 1929). However, the physiological mechanisms by which cortical oscillations affect human behavior are still not fully understood. Some studies indicated that the roles of neural oscillations include feature binding, information transfer mechanisms, and the generation of rhythmic motion output (Klimesch, 1999;Buzsaki, 2004;Canolty et al., 2006;Buzsaki et al., 2013;Iaccarino, 2016). In recent decades, a few breakthroughs have been made by electroencephalography and neuroimaging in this field.
It has been found that neural oscillations play an essential role in the processing of neural information. However, there is little experimental evidence to reveal the function of neural oscillation so far. And it is unlikely to make an accurate explanation for the role of neural oscillation. Further insights into neural oscillations and their relationship to cognitive processes are key research fields of neuroscience. Furthermore, the specific characteristics of the oscillations could be regulated and controlled to achieve functional changes purposefully.
Alpha-band oscillations are the dominant oscillations in the human brain. Although the physiological basis of alpha is still unclear, the origin of it is thought to be related to the and more recent research on alpha-band function has pointed out that alpha-band acts as a significant inhibitory function in information processing of the brain. Previous work pointed out the idea that information is routed by functionally blocking off the taskirrelevant pathways (gating by inhibition). Importantly, this inhibition is reflected by oscillatory activity in the alpha band. And the activation and inhibition of different regions suggested that the parietal lobe was closely related to the attention of visual stimuli and played a crucial role in controlling and monitoring.
Neurofeedback training as an operant conditioning method that helps subjects to control or change their brain activity (Schafer and Moore, 2011). The training enables the modulation of neural oscillations voluntarily by real-time feedback of the brain's electrical parameters. Neurofeedback training is regarded as a safer and more stable method that could control or change neural oscillations than electromagnetic stimulation. Since then it has applied to several application scenarios (Zoefel et  At present, NFT is also applied to improve the performance of various types of BCI. As a form of classical BCI paradigm, SSVEP is also an essential part of the BCI with many applications. SSVEP can be understood as a phase-locked human brain that spontaneously responds to stimulation, and uses signal-to-noise ratio (SNR) to evaluate the signal quality most commonly. Despite decades of research, the performance of SSVEP-based BCI still highly relies on subject and algorithms (Neuper and Pfurtscheller, 2010). SSVEP signal and performance could be influenced by NFT so that Mayra pointed out that subjects can not accept any form of NFT before data collecting in the recent study about SSVEP-based BCI algorithm (Mayra et al. 2018). The study of Ordikhani-Seyedlar suggested that the power of the SSVEP signal was regulated by the attention of various degrees whatever covert or overt attention (Ordikhani-Seyedlar, 2014). Also, Collura designed experiments to demonstrate the relationship between periodic SSVEP and short-term attention (Collura and Thomas, 2002). Moreover, their works pointed out that the stimulation which combining attention and SSVEP was itself a kind of neurofeedback and was expected to be a way to train attention deficits. As for the application of neurofeedback in SSVEP field, Bruno's research claims to improve the visual and auditory cognitive by binaural audio and SSVEP neurofeedback (Bruno et al., 2017).
When it comes to the relationship between attention and SSVEP, Morgan provided more direct evidence. That is, he had asked the subjects to gaze at the center of the screen and set flicker stimulation blocks of different frequencies on both the left and right sides (Morgan et al., 1996). This attention-biased experiment found that the EEG of the subject experienced the frequency of stimulation in the lateral direction of the attention, thus demonstrating the correlation between attention and SSVEP.
Based on the above descriptions, since alpha oscillation has inhibitory function physiologically and SSVEP-based BCI performance seems like high attention required task, we came up with our hypothesis that improving the alpha amplitude of the parietal lobe could increase attention and the SNR of the SSVEP signal to enhance the SSVEP performance.

Design
The flow chart ( Fig. 1) showed the experiment design detailedly. For each subject of the NFT group, the experiment consisted of twenty-four training sessions within two months.
Participants were supposed to have three sessions (every session lasted for an hour) a week. In each session, subjects were asked to play an EEG feedback game. They were free to select three days to receive training in a week, with at least one day between two sessions. In order to ensure the interests of the subjects, the feedback interface was presented to the subjects in the gamification approach with a controllable process. The SSVEP typing test was arranged to compare performance about ten days before and after the two months of training.

Participants
A total of forty-seven subjects (they were randomly assigned to two groups), twenty-five in the NFT group and twenty-two in the control group, took part in the experiment. Subjects with normal vision or corrected vision and no epilepsy or other diseases were selected. The written consent form was received from all of the participants' guardians.
Moreover, the project is also reviewed by the Ethics Committee of the Tsinghua University School of Medicine. The final sample consisted of 24 subjects in the NFT group(six females and eighteen males, 9.42 ± 2.58 years) and 22 subjects in the control group(eight females and fourteen males, 9.86 ± 3.14 years). As van Praag's experiment suggested, aging causes changes in the hippocampus that may lead to cognition and neural plasticity decline in older adults (van Praag, 2005). Therefore, this research chose children to participate in neurofeedback training out of consideration for the young brains are more neuroplastic.
A portable wireless EEG amplifier (NeuSen.W64, Neuracle, China) was used for data recording at a sampling rate of 1000 Hz. Electrode impedances were kept below 10 kOhm for all electrodes throughout the experiment. The experiment was carried out in a natural office environment without any electrical shielding.

alpha-band power ratio
In this study, neurofeedback training was performed to increase the alpha-band amplitude. The alpha-band power was calculated in order to examine the training outcome. Since the absolute value of the alpha-band amplitude is affected by the current state of the subject and the environment, they are not comparable between different trials. In the aim to contrast the results of NFT parallelly, we calculate the ratio of the power which the alpha-band to the 0.8 Hz-20 Hz band to characterize whether the alpha amplitude is improved. In the following formula, X 1 and X 2, denoting the power of the alpha band and full band. In the formula, X denotes the EEG signal and f denotes the frequency.
The resting-state data was obtained by two-minute open-eye and two-minute closed-eye experiments. The closed-eye data totaled 120,000 sampling points with a sampling rate of 1000, and the frequency domain resolution was 1/12 Hz.

SSVEP stimulation
The processing system consists of a stimulation device, an EEG acquisition device, and a processing computer. The stimulation device is ASUS MG279Q 27-inch display with 2560*1440 resolution and 143 Hz refresh rate. The processing device is an Intel i9-9900K@3.6 GHz CPU and DDR4 32 GB memory. Stimulation and processing programs were developed under Matlab 2015a and Psychophysics Toolbox Version 3.
During the experiment, the stimulation was presented in the form of a multi-target flicker.
A start trigger and an end trigger signal were sent to the EEG collector at the beginning and end of each trial. The processing device receives the EEG data from the EEG amplifier in real-time and feeds the judgment results back to the stimulation device after the identification is completed.
As the Fig. 2 and Fig. 3 shown, in the SSVEP test, a 40-target SSVEP speller based on frequency phase modulation coding (frequency range: 8-15.8 Hz, frequency interval: 0.2 Hz) was used . A total of 40 stimulus target blocks were set on the screen, which respectively represents A-Z with a total of 26 letters, 0-9 total ten digits, spaces, commas, periods, and backspace keys. In the experiment, each stimulus target In the entire SSVEP task, in order to guarantee the actual state of subjects and the validity of the data, the staff required each subject to complete six blocks (40 trials each block). Participants can choose to rest between blocks to perform their best. Thus, the entire test takes about 40 minutes. The average SNR of nine leads in the occipital lobe (PO3/4, PO5/6 PO7/8, Oz, O1/2) and the information transfer rate (ITR) were examined to estimate the SSVEP performance of the subjects. In the formulas below, N is total target numbers, A means accuracy, T represents the time required to identify.
For the SNR calculation, flexible band definitions around signals and noise could be made as needed in the engineering project. In this study, SNR was used as an indicator of SSVEP performance to minimize the effect of the background EEG activity across subjects.
Therefore, the power of the signal is specified that the power of a narrow frequency interval centered on the stimulus frequency (f±0.2). At the same time, the power of a broader scale (f±1) is noise. In the formula, X denotes the EEG signal and f denotes the frequency.

Brain network analysis
The whole-brain could be evolved into a complex network while a neuron or a group of nerve cells and the neural connection is regarded as a node and an edge, respectively.
Human response to stimulation and emotional behavior might represent the performance of network overlays. Therefore, the research methods for complex networks could be applied to the research of brain science, and the results of network analysis reflect the structural or functional connection between various brain regions that could provide more clues for explaining the deep mechanism. Traditional network analysis methods are canonical correlation analysis (CCA) and directed transfer function (DTF). In this study, the Referring to Yan's work (Yan and Gao, 2011), the information flow gain ρ was also calculated in this study. The information flow gain ρ is defined as the ratio of the outflow information and the inflow information. The higher the value of the node, the more significant the contribution of the node in the transmission of network information. The smaller the value of ρ, the lower the contribution of the node in the transfer of network information.

Feedback procedure
This study analyses the global brain information connection network induced by SSVEP using the information flow method. In Fig. 4, it can be seen that there is abundant information flow in the SSVEP state network. The parietal-central (FC1, FCz, FC2, C1, Cz, C2, CP1, CP2) plays an essential role in the induction network while the node size stands for ρ. The higher the value ρ of the node, the more significant the contribution of the node in the transmission of network information. Consistent with this, in 2000 Corbetta reported the control effect of the parietal lobe in visual tasks (Corbetta, 2000). Taken together, it suggested that the parietal lobe plays as a central region in visual tasks with attention reorienting effect. The position of the neurofeedback training was selected in the parietal lobe, in which the location with the lowest alpha amplitude was regulated. The electrode selected varies from subject to subject. The training only targets one electrode-position in order to better detect the result.
The NFT system is arranged like Fig. 5. Considering that suspension will require the neuroplasticity of children is stronger, the study selected subjects aged from 7 to 12. In order to attract the interest of the subjects and make them be co-operative, the neurofeedback interactive interface is designed in the form of a game (the interface display like Fig. 6). The program has three options and subjects could select one of them to execute: a. Pacman (A kid eats beans on a given route. Feedback parameters above and below the threshold set corresponding to game pause and progress status respectively). b.
Fly the airplane (The aircraft flies on a given route. Feedback parameters above and below the threshold set corresponding to game pause and progress status respectively). c.
Racing car (Driving the car on a given route. Feedback parameters above and below the threshold set corresponding to game pause and progress status respectively). Regardless of the form of the game, the subject is always able to keep the EEG parameters above the feedback parameters, thus achieving the effect of neurofeedback training.
The setting of the feedback parameter is divided into two parts, first is the frequency setting and the other is the amplitude setting. Before the formal training, the subjects tested for different feedback frequencies in alpha-band. And the frequency under which the rate of success (neurofeedback training) achieves 60% − 70% is set as the feedback frequency. For the amplitude, it has been found that the amplitude is set to 0.2-0.3 times of the alpha amplitude when subjects are closing their eyes. In this case, the success rate of the training is also 60% − 70%. The feedback process is controlled as Fig. 7 shown.
Besides, we adjusted the feedback parameters appropriately according to the success rate (usually increasing the feedback frequency) as the training goes on. The sampling rate resampling to 250 Hz during calculation while the original sampling rate is 1000 Hz. The length of the calculation window is 2 seconds, and the overlap of the calculation window is 50%. Data update time shall not exceed 1 s.
There are 2000 points used to compute the feedback parameters. The sampling rate is resampled to 250 Hz during calculation while the original sampling rate is 1000 Hz. The length of the calculation window is 2 seconds, and the overlap of the calculation window is 50%. Data update time shall not exceed 1 s. The welch method is used for the power computation. And there is a bandpass filter (0.5-40 Hz) added to the original data.
Subjects involved in neurofeedback training will sit in front of the computer wearing an EEG cap, gaze at the screen and follow the instructions to complete the task. The system can feedback EEG parameters in real-time and determine whether the respondents are successful. If the feedback is successful (feedback parameters above the threshold), subjects can continue the game. Otherwise, the suspension will require the subject to adjust their status. At this time, the instructor would lead the subject to pay attention to the game. And the subjects should try to adjust state by themselves.

Integrated Visual and Auditory Continuous Performance Test
The Integrated Visual and Auditory (IVA) continuous performance test (CPT) is a screening tool testing sustained attention (Sandford and Turner, 2000). It was developed by John Sandford (psychologist) and Anne Turner (physician). The initial test consisted of a series of pseudo-randomized auditory and visual stimulation (totally 500 trials) and asked the subject to click on the target they heard or saw. The entire test lasted approximately 20 minutes and our research selected part of the parameters to examine subjects. They were asked to complete some paper quality tables and tested under the guidance of the instructor. The test is a real-time task consisting of repeated auditory and visual stimulations. During the test, the instructor would guide the subject to observe a series of graphics and texts on the scale. Subjects would also listen to a short essay with overt information. The instructor would then ask the participant to give the information required and answer the questions on the scale. The test is used to monitor the effectiveness of neurofeedback training or medication generally.

Neurofeedback Training results
We recruited twenty-four participants, each participant completed 24 training sessions in two months (24 hours in total count). Among these participants, 91.7% were successfully trained (22 of 24). Successfully trained means the alpha-band power of subjects is higher after training. In the Fig. 8, the average alpha-band power ratio of the 24 subjects before and after training is 32.93 and 38.58, respectively. Thus the training brings about 17.16% improvement in the alpha-band power ratio, and the results show a significant difference (t = 23,p = 0.00015) by a paired-sample t-test. . Kim's study tested 100 healthy children on the same standard behavioral tests, and we mapped their scores(the original scores range 0-155) to 0-10 for comparison. The statistical results are shown in the following table, each of which is scored from 0-10, with 10 being the best. As the table.1 shown, the mean value of 7 factors before training is 5.33 ± 0.97, while the index after training is 6.76 ± 0.93, which shows a significant difference in a paired-sample t-test (t = 23, p = 0.00172). It indicates that the training brings impressive improvements to visual and auditory attention.

SSVEP results
Improving SSVEP performance is one of the ultimate goals of this work, and it is also the result supposed to be observed after neurofeedback training. After statistical analysis, the results showed that 87.5% of the subjects (21 of 24) had an improvement with various degrees.

Accuracy results
Each participant took part in an SSVEP test of 240 trials within 6 blocks. The average accuracy of 6 blocks was calculated as BCI accuracy (Fig. 9). The results showed that 79.17% of the subjects (19 of 24) improved in some way, while the other five subjects did not improve as they achieved a relatively high accuracy rate (about 90%) in the first test.
The average accuracy before training is 78.9% and the figure after training is 83.6%, which shows a significant difference(t = 23, p = 0.0000049). As for the control group, the average accuracy before training is 79% and the figure after training is 79.4%, with no significant differences.

ITR results
According to the formula, the average ITR of SSVEP of the 24 participants was 103. After training, the average score of SSVEP of 24 subjects was 119, bringing about 15.6% improvement (see the Fig. 10). After a paired-sample t-test, there was a significant difference (t = 23, p = 0.0494). And for the control group, the average ITR after two months is only 103, which is 112 before.
It is noteworthy that three of the subjects do not display an increase of SSVEP ITR because they achieved a high level in the first test, and the performance index declines after training are within a normal fluctuation range.

SNR results
SNR is widely used in the engineering field as an indicator to measure the quality of the We also analyze the SSVEP frequency response and select the response of 8,9,10,11,12,13,14,15 Hz for all the trials to calculate the topographic map (see Fig. 13). In the SSVEP frequency response topographic map, the experimental group and the control group responded similarly in the first test. However, in the second test, the experimental group have less interference and stronger signal.
We performed a paired t-test on the SSVEP frequency response before and after two months. The p-value obtained was plotted into a statistical topographic map, where the scales range is from 0.2 to 0. That is, the darker region has smaller p-value, indicating that the difference is larger. Vice versa, the lighter has bigger p-value. Above all, these are the locations which have differences: AF3, FC4, CP1, Pz, O2. Notably, the O2 has a significant difference for the experimental group. These results show that the training brings differences to the parietal lobe, the occipital lobe, and the frontal lobe.
Furthermore, changes in the occipital region imply that the interference there is suppressed and the SNR increased. The change in the parietal lobe may be related to the enhancement of the control network via neurofeedback regulation. As for the frontal lobe, the difference may be due to the more concentrated attention of the subjects and the less disruptive interference.

network analysis results
Based on the statistical map of the frequency response difference of the two SSVEP tests, it can be seen that there is a significant difference in the right parietal lobe, occipital region and frontal lobed, so the three locations AF3, FC4, CP1, O2 are selected as the region of interest to analyze the cortex (result shown in Fig. 14 & Fig. 15). Overall, in the connected network of the SSVEP task, the parietal lobe is a hub of information transformation, and there is a large amount of information flowing from the parietal lobe to the occipital lobe and frontal lobe. This result is also consistent with the previous studies (Yan and Gao, 2011). After calculation, the information of each node of the test group before and after the test was compiled as shown in Table 2 and Table 3, and the flow of information is from row to column. Table 4 suggested that the parietal lobe node CP1 contributes the most in network information transmission. And according to Table 2 and Table 3, the training strengthened the connection of the parietal lobe node to O2 and FC4.
As the information flows from the parietal lobe to the occipital lobe, we further compute the alpha power of the occipital lobe. As shown in the above figures, the occipital lobe alpha power in the task-state was significantly reduced after the neurofeedback training of up-regulating the parietal lobe alpha, while the control group remains normal. The decrease of background alpha power in the occipital lobe reduces the frequency-domain correlation between signal and background noise, which is also consistent with our SNR analysis results.

Discussion
In accordance with the hypotheses, the trainability of up-regulating NFT in the parietal lobe has been confirmed. The results proved that the training method of this study was conducive to reinforcing the control network and thus enhanced the ability to complete Another work was based on alpha down-regulating neurofeedback training to improving the SSVEP performance (Wan et al., 2016). Since Wan's work seems to conflict with the work of this study, it is worthwhile to discuss the difference between the two works in detail. In the project of Wan, a two-day alpha down-regulating feedback neurofeedback training on the occipital region was conducted on 20 subjects, and it was found that this between our experiment and Wan's. This study focuses on the parietal lobe while the occipital lobe was regulated in Wan's research -however, two research access to the same goal that enhances the SSVEP-based BCI performance through different aspects.
In the SSVEP test of Wan,7.05,7.5,8,8.57,9.23,10,10.9,12,13.33 Since that, the BCI accuracy is much dependent on whether the frequency response is precise. As is known to all, the alpha-band is usually defined as 8 In fact, the human brain is not merely a passive response to the stimulus, but a top-down processing mechanism with macro-control. In this mechanism, the flow of information is

More description about controls and functional
Prior research has reported that the control network and functional network may exist and provided a lot of direct or indirect evidence. Foxe and Bauer recommend models implicating parieto-occipital areas prominently in the directing and maintenance of visual attention (Foxe et al., 1998;Bauer, 2006). As well as, the results indicated that the network of control and functional might not be wholly separated. In the corresponding difference analysis of SSVEP data before and after training (see Fig. 10), it can be seen that neurofeedback training affects the parietal lobe, occipital lobe, and frontal lobe. The alteration takes place in the occipital lobe and the frontal lobe is due to the increase of attention and the decrease of signal noise. Simultaneously, the differences observed in the parietal lobe advocates that the NFT enhances the capacity of the control network. In Srinivasan's fMRI studies (Srinivasan, 2007), it was also found that some of the occipital voxels were positively correlated to the frontal voxels forming a large-scale functional network when visual input. Coupled with the parietal lobe push forward an immense influence on regulating of visual stimuli, we could suggest that the human brain responds to visual stimuli by arousing a parietal-occipital-forehead control/ functional network.
Moreover, parietal-occipital-forehead region work as an inseparable system.
The outcome of network connection analysis also verifies the inference that the frontal lobe, parietal lobe and occipital lobe work in a network within the SSVEP visual task. The parietal lobe serves as the sole hub of information output and transmission. In Table 2 and Table 3, it can be seen that the information output of the parietal lobe (CP1 to O2, CP1 to FC4) is increased after training in the NFT group, which also confirm the enhancement of the control capability of the parietal lobe.

Worth and innovation
In our work, a total of 24 children have arranged a neurofeedback training for two months.
After investigations, we found that 1-2 months seems to be the shortest period for NFT which could produce a stable effect. Some previous research suggests that the resting of the keys to realizing brain-computer interaction in the future. In order to realize braincomputer interaction and the enhancement of human brain function, stable and sustainable regulating is crucial. Short-term and unstable modulating methods may bring some unexpected trouble.
Despite people's extreme desire to know more about the human brain, breakthrough in brain science is driven by strictly prohibit prudence experiments. This also requires a controllable and stable experimental design to avoid any possible damage and risk. The choice of children subjects in this study undoubtedly increases the difficulty of research.
Nonetheless, take the young brains are more neuroplasticity into consideration, such an attempt makes sense. In most studies on neurofeedback training, the choice of subjects is adults (Angelakis et al., 2007;Zotev et al., 2014), and research on children is relatively rare.

Conclusions
In this work, we found a kind of neurofeedback training method and got the following

Consent for publication
Consent to publish was obtained from their parent or legal guardian.

Availability of supporting data
The raw/processed data required to reproduce these findings cannot be shared at this time as the data also forms part of an ongoing study. We could consider publishing some valuable data in case of acceptance.
The authors declare that they have no competing interests.  Protocol description of pre-and post-training for the NFT group and non-NFT group. The entire experimental process includes data acquisitions, feedback training, and data processing. The control group and the experimental group are only different from whether or not feedback training was performed.

Figure 2
The SSVEP stimulation interface consists of 40 stimulus target blocks on the screen, which respectively represent A-Z with a total of 26 letters, 0-9 total ten digits, spaces, commas, periods, and backspace keys.  The figure shows the global information flow of functional connectivity in the SSVEP task. In figure A, the node size is represented by ρ and the region is distinguished by color. It can be seen that the eight leads of the parietal lobe play a vital role in the information network throughout the network. Figure B shows the location of the core nodes of the parietal lobe on the cortex.  The neurofeedback training process is shown in the figure. The real-time EEG parameters of the subject are collected and compared with the set threshold. If the EEG parameter is above the threshold, the game continues. Otherwise, the game suspends, and the subject should adjust the state as required. The entire process circulates to finally achieve the training goal. (Error bar indicates the SD, * represents the significant difference, p < 0.05). SSVEP-based BCI accuracy is dramatically increased after neurofeedback training.
The figure shows the distribution of the SSVEP-BCI accuracy. Box graph statistical recognition accuracy distribution makes it clear that training makes the classification accuracy improve dramatically. And there were significant differences in the results before and after training of the experimental group, but almost no differences in the control group. It can also be seen from the diagram that the distribution of the first test results of the experimental group is approximately the same as that of the control group, in which the singular points 43 are not in the statistical range and have been marked. (Error bar indicates the SD, * represents the significant difference, p < 0.05).

Figure 10
The NFT speeds ITR of the SSVEP system (50% chance of resampling to eliminate the singularity repeat for 100 times). The blue dots indicate the results of the experimental group and the green dots indicate the ITR of the control group. The red line is the reference line. The figure shows that most blue dots tend to be distributed in the upper left area, which means that the training results are better than that before training. While the green dots tend to remain at the same level, distributing around the red line.

Figure 11
Statistical averaged SNR of SSVEP signal is considerably improved after NFT in the experimental group. It can be seen that the SNR of the first test of the training group and the test of the control group is not much different, while the second test of the experimental group showed a significant improvement in signal to noise ratio. This is a phenomenon not observed in the control group, excluded the possibility that the skilled use of the SSVEP-based BCI system is the main reason for the improvement of the experimental group. (Error bar indicates the SD, * represents the significant difference, p < 0.05).

Figure 12
The figure shows the SNR changes in each frequency band of the experimental group and the control group. The experimental group had improved in each frequency band, while the change in the control group seems like randomly.

Figure 13
SSVEP frequency response and paired-sample t-test P-value statistical topographic map. The graphic color of the second data acquisition of the experimental group indicates that the training has achieved the purpose of reducing interference. For the experimental group, these are the locations that have differences (p<0.1): AF3, FC4, CP1, Pz, O2. Importantly, the O2 has a significant difference (p<0.05).

Figure 14
The figure shows the information flow of cortical functional connectivity in the SSVEP task among the right parietal lobe, occipital region, and frontal lobe. The information flow calculated using the directed transfer function (DTF).

Figure 15
The picture shows the detailed brain information flow of two groups of subjects in the SSVEP-based task state. The grid elements of the Matrix are 59*59 electrode leads, and the subareas have been marked. The flow of information in the graph is from the row to the column. It can be seen that almost all the main hubs in the network have the flow of information to the whole brain, while the unimportant nodes have little information exchange.

Figure 16
The picture shows the background alpha component (task-state) of the occipital lobe. In figure A, the alpha power of the experimental group in the occipital lobe decreased significantly, and the results passed the paired-sample t-test (t=23, p=0.0461). In figure B, the alpha power of the control group in the occipital lobe shows no statistical regularity. And the average and variance value of two groups before and after the experiment is given in figure C.