Optimizing Visual Stimulation Paradigms for User-Friendly SSVEP-Based BCIs

In steady-state visual evoked potential (SSVEP)-based brain-computer interface (BCI) systems, traditional flickering stimulation patterns face challenges in achieving a trade-off in both BCI performance and visual comfort across various frequency bands. To investigate the optimal stimulation paradigms with high performance and high comfort for each frequency band, this study systematically compared the characteristics of SSVEP and user experience of different stimulation paradigms with a wide stimulation frequency range of 1–60 Hz. The findings suggest that, for a better balance between system performance and user experience, ON and OFF grid stimuli with a Weber contrast of 50% can be utilized as alternatives to traditional flickering stimulation paradigms in the frequency band of 1–25 Hz. In the 25–35 Hz range, uniform flicker stimuli with the same 50% contrast are more suitable. In the higher frequency band, traditional uniform flicker stimuli with a high 300% contrast are preferred. These results are significant for developing high performance and user-friendly SSVEP-based BCI systems.


I. INTRODUCTION
B RAIN-COMPUTER interface (BCI) is an advanced tech- nology that establishes a direct connection between the human brain and external devices [1], [2], [3].Among various BCI paradigms, the steady-state visual evoked potential (SSVEP) [4] is widely utilized due to its advantages of low training demand and high information transfer rate (ITR).SSVEP is a neural response of the brain to continuous periodic visual stimuli with time and phase locking characteristics [5], [6], [7].
The response intensity of SSVEP is highly related to the stimulation frequency [8], [9].According to the relationship between stimulation frequency and SSVEP amplitude, SSVEPs can be divided into three frequency bands: low (4-12 Hz), medium (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and high (>30 Hz), with a peak value appearing in the amplitude spectrum in each band [10], [11].As the stimulation frequency increases, the amplitude of the fundamental frequency gradually decreases, and the high frequency band has the lowest amplitude among the three frequency bands [12].Therefore, in many SSVEPbased BCIs with high ITRs, the stimulation frequencies were commonly chosen in the lower and middle frequency bands.For example, Nakanishi et al. designed an online SSVEP-based BCI system with 40 targets, employing a frequency and phase hybrid encoding method in 8-15.8Hz, achieving a maximum ITR of 325.33 bits/min [6].But in fact, high frequency stimulations offer substantial advantages in enhancing user experience, which is also an essential condition for the practical application of BCI systems.Jiang et al. constructed a four-target SSVEP-based BCI system at 60 Hz, and the participants reported no perception of flickering during the experiment [13].However, owing to the weak response, improving the ITR of high-frequency SSVEP BCIs remains a challenging task.
Some studies have made attempts to simultaneously optimize the performance and user comfort of SSVEP-based BCI systems.Lee et al. effectively reduced subjective visual discomfort by optimizing the duty cycle within the flicker cycle, and achieved an ITR of 25.08 bits/min using six phase encoded targets at a stimulus frequency of 13.16 Hz [14].Chang et al. designed a six-target BCI system based on amplitude-modulation visual stimulation, attaining an average ITR of 30.4 bits/min in a state of slight visual fatigue [15].Liu et al. employed an advanced refresh technology and a novel stimulation presentation method to achieve 40-target encoding under low flickering perception conditions.The maximum online ITR reached 63.86 bits/min in the frequency band of 55-62.8Hz [16].Zhao et al. designed a 40-target BCI system operating in the frequency band of 8 to 15.6 Hz.They utilized peripheral vision to minimize the total number of stimuli, ultimately achieving an ITR of 109.2 bits/min and effectively enhancing the comfort of the subjects [17].
In addition to optimizing stimulation coding methods, there are also studies exploring the optimization of stimulation properties.Traditional uniform flicker stimuli flash between the brightest (white) and darkest (black) luminance levels [5], [18], [19].Ladouce et al. achieved high classification performance and enhanced user comfort by reducing the amplitude depth (i.e., the maximal luminance and the contrast between the two alternating states of the stimuli) of uniform flicker stimuli in the 13-24 Hz frequency band [20].Waytowich et al. discovered that a checkerboard stimulation paradigm with a spatial frequency of 2.4 c/ • can maximize ITR while reducing subjective visual irritation.They attained a theoretical ITR of 45.3 bits/min using four-target encoding in a frequency range of 6-8.57Hz [21].Ming et al. introduced an innovative checkerboard-like stimulation pattern composed of a 50% stimulus region and a 50% background region and achieved a balance between performance and user experience by reducing the proportion of flickering areas and optimizing spatial contrast [10].In both low frequency (11-15 Hz) and high frequency (38)(39)(40)(41)(42) conditions, the online ITR attained 124.0 bits/min and 109.0 bits/min, respectively, when utilizing the black-background checkerboard-like stimulation paradigm.
Recently, a new grid stimulation pattern has been introduced [22].The grid stimulation pattern utilized an approximately 25% stimulation area (i.e., flickering grid cells are evenly spaced over a stationary background) and, after optimizing spatial contrast, achieved comparable classification performance in low and medium frequency bands as traditional uniform flicker stimuli, while also delivering a satisfactory user experience.Compared to the traditional uniform flicker stimuli, the grid stimuli not only differ in the proportion of stimulation but, more significantly, they convey spatial information of positive or negative contrast due to the presence of a background region with constant luminance.When recognizing objects, ON bipolar cells in the retina respond to visual signals brighter than the background, while OFF bipolar cells respond to signals darker than the background [23].Therefore, the ON and OFF grid stimuli represent visual information with positive and negative contrasts, respectively, activating independent parallel ON and OFF visual pathways, respectively [24].Previous investigations have documented the differences between the ON and OFF visual pathways.The neurophysiological research on macaque monkeys found that the number of ON and OFF neurons in the primary visual cortex (V1) was not equal, and the OFF neurons were dominant in layers 2/3 of V1 [25].Similarly, in visual perception research focusing on light and darkness, human observers detected lights more slowly than darks, with a 6-14 ms time difference [26].In terms of visual evoked potentials (VEPs), Zemon et al. found that the OFF subsystem had finer spatial tuning and greater contrast gain than the ON subsystem, and stimulus attributes such as frequency, spatial frequency, and contrast can affect its VEP response [27].Ming et al. observed that under high frequency (38)(39)(40)(41)(42) stimulation conditions, the classification performance of black-background checkerboardlike stimulation (ON-type stimulation) was higher than that of white-background checkerboard-like stimulation (OFF-type stimulation) [10].Interestingly, stimuli presented in the low and medium frequency bands demonstrated a noteworthy advantage in achieving a balance between performance and user experience when utilizing OFF grid stimulation with a 50% contrast, as opposed to ON grid stimuli [22].In summary, the positive and negative contrast grid stimuli have been proven to be feasible for stimulation paradigms in SSVEPbased BCIs.The limited stimulation area and low contrast properties provide great advantages and potential for building user-friendly BCI systems.
The purpose of this study is to optimize the stimulation paradigms that employ grid stimuli and uniform flicker stimuli under different stimulation frequencies to enhance the performance and user experience of SSVEP-based BCIs.The frequency scanning experiment studied the effects of stimulation frequency (1-60 Hz with an interval of 1 Hz) on the SSVEP characteristics and user experience of grid stimuli and uniform flicker stimuli at 50% contrast, and compared them with traditional uniform flicker stimuli with 300% contrast.Based on the comparison results, the optimal stimulation paradigms of low, medium, and high frequency bands were selected to construct four-target BCI systems.The results of the BCI experiment including the user experience scores and overall preference rankings for each BCI system in different frequency bands were analyzed.The findings of this study could provide references for the design of high-performance and user-friendly BCI systems.

A. Subjects
Fifteen healthy subjects (eight females) participated in this study.The age of the participants varied from 24 to 29 years, with an average age of 26.5 years.All participants possessed regular or corrected-to-normal vision.Before the experiment, participants were guided through the experimental procedures and were explicitly informed of their right to withdraw from the experiment at any time.They signed the informed consent forms approved by the Institution Review Board of Tsinghua University (NO.20230058).

B. Stimulation Scheme
Fig. 1 illustrates four stimulus paradigms utilized in the experiment: ON grid, OFF grid, and uniform flicker stimulation paradigms, all with the same low contrast values, along with the high-contrast uniform flicker stimulation pattern used for comparison.In the whole pattern, there is a constant luminance gray background area and a flickering area (i.e., the uniformly distributed grid cells) where luminance changes over time.During stimulation, the luminance of ON grids maintains above the background luminance, while the luminance of OFF grids maintains below the background luminance.The value of spatial (luminance) contrast can be determined through the following definition of the Weber contrast [28]: where L c is the highest or lowest luminance of the grid cells during the stimulation period, and L b is the luminance of the background area.All luminance values in this experiment were measured by the screen luminance meter (Sanpo SM208).
In the experiment, the luminance of the background region was 52.3 cd/m 2 .For the ON grids, L c represented the highest luminance at 78.6 cd/m 2 , while for the OFF grids, L c represented the lowest luminance at 26.3 cd/m 2 .The spatial contrast value for both the ON grid and the OFF grid was 50%.Similarly, a uniform flicker stimulus with the same contrast varies in luminance between the maximum luminance of the ON grids and the minimum luminance of the OFF grids.Thus, the contrast of the uniform flicker stimulus that flashed between 26.0 and 78.6 cd/m 2 was determined as 50%, and the contrast of the uniform flicker stimulus that flashed between 0 and 400 cd/m 2 was 300%.An individual target stimulus measures 310×310 pixels (7.10 • ), and a single grid cell has a size of 10×10 pixels.The spatial frequency for both the ON and OFF grid stimuli was determined as 2.17 c/ • , guided by the optimization results from the checkerboard-like stimuli in [22].
According to the results of the frequency scanning experiment within 1-60 Hz, the primary stimulation paradigms for various frequency bands were selected to construct high-performance and high-comfort BCI systems.Stimulation frequencies of 2, 5, 10, 20, 30, and 60 Hz were chosen to establish four-target BCI online systems with phase encoding.The stimulation frequencies of the four targets were the same in one distinct BCI system.Specially, stimulation at frequencies of 2, 5, 10, and 20 Hz utilized the ON 50% and OFF 50% stimulation patterns, while the stimulus of Flicker 50% was employed at 30 Hz and the stimulus of Flicker 300% was used at 60 Hz.Fig. 2 shows the user interfaces and four-target encoding scheme corresponding to the four stimulation paradigms.

C. Stimulus Presentation and Data Acquisition
The experiment employed a 24.5-inch LCD display (Alienware AW2521H) with a resolution of 1920 ×1080 pixels and a screen refresh frequency of 240 Hz to present the visual stimulation.Every stimulus driven by a sampled sine wave was implemented using the Psychophysics Toolbox Version 3 (PTB-3) within the MATLAB environment [29].For the stimulus frequency scanning experiment, EEG data from 64 channels with a sampling rate of 1000 Hz were recorded using a Neuroscan Synamps2 system.Electrode placement followed the international 10/20 system.The reference electrode was positioned at the vertex, and electrode impedance was consistently maintained below 20 k .For the four-target BCI experiment, EEG data of 9 channels (Pz, PO3, PO4, PO5, PO6, POz, Oz, O1, and O2) were recorded at a sampling rate of 20000 Hz.The high sampling rate ensures a high phase resolution of SSVEPs when using phase encoding for target identification.Event triggers, generated by the stimulus presentation program, were conveyed to the amplifier through a parallel port.A built-in 50 Hz notch filter was used to eliminate power line noise, and a band-pass filter was set from 0.1 to 100 Hz to preserve the wideband spectral characteristics of EEG signals.Before the experiment, a comprehensive assessment of the monitor's refresh rate was conducted to guarantee the stability of the stimulus presentation [18].

D. Experimental Procedure
In the frequency scanning experiment, a comprehensive comparison of 240 different stimuli was conducted.These stimuli comprised ON 50%, OFF 50% grid patterns, along with Flicker 50%, and Flicker 300% uniform flicker patterns, covering a frequency range from 1 to 60 Hz with a 1 Hz Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
increment.The entire experiment was conducted in a well-lit room with an average luminance of 23 cd/m 2 .Environmental brightness was maintained consistent by keeping the overhead LED lights on and using blackout curtains to control external light exposure.The participants were situated approximately 70 cm away from the monitor.Due to the wide variety of stimulation patterns, the experiment was divided into four sessions, with each participant completing all sessions on different days.15 different stimulus frequencies were conducted in one session, and the stimulus frequencies of the four sessions set as follows: 1) 1-57 Hz with an at interval of 4 Hz for the first session; 2) 2-58 Hz with an at interval of 4 Hz for the second session; 3) 3-59 Hz with an at interval of 4 Hz for the third session; 4) 4-60 Hz with an at interval of 4 Hz for the fourth session.Each session comprised 12 blocks, with each block consisting of 60 trials that corresponded to 60 randomly occurring stimuli for the four conditions (the ON 50%, OFF 50%, Flicker 50%, and Flicker 300%).Each trial lasted for 7 s.Initially, a red cross appeared for the first second, indicating to the subject that the stimulus was about to begin.Subsequently, a 5 s flickering stimulus was presented at the centre of the screen.Simultaneously, the subjects were required to maintain focus and avoid blinking.After the flickering stimulus stopped, the red cross disappeared, and the subjects rested for 1 s.
After collecting all EEG data for four sessions, participants were required to complete a subjective assessment for each stimulation paradigm.A Samsung LC49G95T monitor with a resolution of 5120 × 1440 pixels and a refresh frequency of 240 Hz was used to simultaneously display 60 visual stimuli randomly on the screen.The screen's capability to present more stimuli on a single page facilitated their mutual comparison.To evaluate the 240 stimuli, each participant navigated through four pages on the screen, filling out a questionnaire corresponding to the sequence of stimuli.The questionnaire included three dimensions [10]: 1) Comfort level rating on a five-point scale (1-5: very uncomfortable, uncomfortable, slightly uncomfortable, comfortable, very comfortable).2) Flicker perception rating on a five-point scale (1-5: very annoying, annoying, slightly annoying, perceptible, imperceptible).3) Preference rating on a five-point scale (1-5: very disgusting, disgusting, neutral, likeable, very likeable).In the BCI experiment, ten distinct stimuli were employed (ON 50% and OFF 50% at 2, 5, 10, and 20 Hz, Flicker 50% at 30 Hz, and Flicker 300% at 60 Hz).EEG acquisition was divided into two parts: initially collecting individual EEG templates for each target, and subsequently applying the specific template for real-time target classification tasks.During the training stage, 4 blocks were collected for each stimulus pattern, each block consisting of 20 trials, facilitating five randomized traversals of the four targets.Each trial lasted for 2 s, comprising a 1 s cue (a red cross with a size of 0.55 • appeared) followed by a 1 s stimulus presentation.Participants were required to take a brief break or blink and shift their gaze to the upcoming stimulus target during the 1 s cue period.Throughout the stimulus presentation, participants were instructed to maintain focus, avoiding body movements and blinking.Finally, 20 trials of EEG data were gathered for each target to acquire a training template.During the test stage, 2 blocks were collected for each stimulus pattern.Differing from the training phase, after each target stimulus presentation, participants received real-time feedback based on the classification results from the online data analysis program.If recognized correctly, participants would hear a beep sound; otherwise, the absence of sound indicated an incorrect identification.This feedback period was integrated within the cue duration of the subsequent target.
Following all EEG data collection, participants were first required to assess the user experience with the ten stimulation interfaces.The questionnaire settings were kept consistent with those used in the frequency scanning experiment.After that, a bubble sorting technique was applied to rank the ten BCI systems.Participants were required to rank these ten stimulation interfaces based on their joint consideration of individual recognition accuracy and subjective perceptions.The weighting between accuracy and user experience was subjectively determined by the participants, with examples of possible ratios being 7:3, 1:1, or other proportions.During the evaluation, the classification accuracy (in percentage) was marked in the upper left corner of each user interface.Participants were instructed to review the accuracy initially and then, based on their personal ranking criterion, mark the superior of the two consecutively presented stimuli.This pairwise comparison process was iterated 45 times, enabling the ranking of all interfaces.The goal was to identify the most appropriate user interface that strikes a balance between accuracy and user experience, taking into account the perspective of the participants.

E. Target Classification Algorithm
The filter bank [30] ensemble task-related component analysis (e-TRCA) [6] algorithm was adopted for target recognition in this study.The e-TRCA algorithm extracts task-related components from signals by maximizing the covariance between trials.The filter bank algorithm can further improve classification performance by combining SSVEP components at the fundamental and harmonic frequencies.Firstly, all EEG epochs down-sampled to 250 Hz were segmented into N m sub-band components.The Chebyshev Type I band-pass filter used in the first sub-band was configured with the parameters [u, 90] Hz, where the values of parameter u were determined towards the highest classification performance.While the band-pass filters in the m th sub-band were set to [ f × m − 0.2, 90] Hz to preserve harmonic responses, where f represents the stimulus frequency and m >= 2. Subsequently, the optimal spatial filter w (m) n of the m th sub-band of the n th target can be calculated using the TRCA algorithm.An ensemble spatial filter of the m th sub-band can be acquired as: where N i is the number of targets.Then, an independent template Xn (m) of the m th sub-band of the n th target can be obtained by averaging the training data across all trials.The Pearson correlation coefficient between the m th sub-band test data Z (m) from a single trial and the corresponding template Xn (m) can be calculated as: The feature used for classification is identified as a combined correlation coefficient for all sub-band components: where the weight for each sub-band component was defined as follows: Here, a and b are constants, and their values can be determined either based on prior experiential settings or by employing a grid search method to maximize classification performance.
Ultimately, the label of test data Z can be judged as the target index with the highest correlation coefficient σ n :

F. Performance Evaluation
The EEG topographic analysis utilized 64 channels and the rest of the analysis employed 9 channels (Pz, PO3, PO4, PO5, PO6, POz, Oz, O1, and O2) in the parietal and occipital regions.The fast Fourier transform (FFT) algorithm was adopted to calculate the amplitude of fundamental frequency and harmonics frequencies of SSVEPs.The SNR of SSVEP (in decibels, dB) is calculated as follows: where y ( f ) is the amplitude at the stimulus frequency f , f is the frequency resolution of the amplitude spectrum.In this study, f was 0.2 Hz and the number of adjacent frequencies that were included in the noise was 8. Classification accuracy can be used for evaluating the performance of BCI systems.In the frequency scanning experiment, the optimal frequency range and filter bank parameters for each stimulus paradigm were determined through grid search.In addition, the offline classification accuracy was calculated using the leave-one-out cross-validation method.In the four-target BCI experiment, the same filtering parameters as those used in the frequency scanning experiment were employed, and real-time classification results were obtained by the online data analysis program.Furthermore, ITR can provide a standardized metric for comparing the performance of different BCI systems.The definition of ITR (in bits per min, bpm) is as follows: where M is the number of targets, T represents the average target selection time and P is the classification accuracy.

G. Statistical Analysis
This study employed repeated measures analysis of variance (ANOVA) for data analysis, aiming to assess the differences in BCI performance and user experience across the four different stimulus paradigms.The Greenhouse-Geisser correction was applied when the assumption of sphericity was not satisfied.Subsequently, post hoc pairwise comparisons were conducted, using the Bonferroni correction with a significance level set at 0.05.

A. Signal Characteristic Analysis
Fig. 3 depicts the SSVEP responses of four paradigms at stimulation frequencies of 2, 12, 30, and 45 Hz.Fig. 3(a), (c), and (d) demonstrate that all four paradigms are capable of eliciting robust periodic SSVEP signals at frequencies below 30 Hz.In addition to responses at the fundamental frequency, significant SNR responses are also observed at various harmonic frequencies.For example, the fourth harmonics for grid stimuli and the seventh harmonics for uniform flicker stimuli at 12 Hz both exhibit strong responses.Furthermore, at frequencies above 30 Hz, the uniform flicker stimulation paradigms display a more pronounced response.
Topographic maps in Fig. 3(b) illustrate the amplitude of the fundamental frequency across the four example stimulation frequencies.It is apparent that regions with substantial SSVEP responses are primarily localized in the occipital area of the brain.Specifically, at 2 Hz, the ON 50% and OFF 50% grid stimulations yield stronger fundamental frequency responses compared to the uniform flicker stimulations.At 12 Hz, the activation regions of ON 50%, OFF 50%, and Flicker 300% are similar in size, while Flicker 50% exhibits a minimal response.As the stimulation frequency increases to 30 Hz and 45 Hz, the responses of ON 50% and OFF 50% grid stimulations are the weakest, while Flicker 50% shows a higher response than the grid stimulations, and the responses of Flicker 300% are the highest.
The amplitude and SNR of SSVEPs at the fundamental frequency and second harmonic for the four stimulus patterns are depicted as a function of stimulation frequency in Fig. 4. As the stimulation frequency increases, there is an initial rise followed by a decline in the overall amplitude response, with variations in the inflection points among the four stimulation paradigms.There is a significant double-peak feature observed in the fundamental frequency amplitude response curves of ON 50%, OFF 50%, and Flicker 300% stimulation paradigms in Fig. 4(a).The two peaks of grid stimuli appear at 4 Hz and 14 Hz, while the two peaks of uniform flicker stimuli appear at 12 Hz and 22 Hz.Before 17 Hz, the amplitude responses of grid stimuli are the leading among various stimulus paradigms, and after that the traditional Flicker 300% has the strongest response.Similar to the amplitude response, the SNR response of grid stimuli is stronger than that of uniform flicker stimuli before 18 Hz, but the trend is reversed after 18 Hz, and the curves of ON 50% and OFF 50% remain closely matched.From the perspective of the second harmonic response, in the frequency range above 8 Hz, the response of uniform flicker stimuli is stronger than that of grid stimuli (amplitude in Fig. 4(c) and SNR in Fig. 4(d)).Through the analysis of Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.SSVEP signal characteristics, the grid stimulation shows a significant advantage in the low and medium frequency bands.

B. Performance Comparison Between Stimulus Paradigms
To further assess the performance of the four stimulation paradigms when applied in a BCI system, a four-target classification task using phase encoding [13], [31], [32] under each stimulation frequency was designed for each stimulation paradigm.Using different time shifts, four data segments with initial phase values of 0 • , 90 • , 180 • , and 270 • and a duration of 1 s were extracted for classification.Data for the first 140 ms of each trial were removed from the analysis to prevent the impact of the transient event-related potential [19].The filter bank e-TRCA algorithm and the leave-one-out cross-validation were adopted to estimate the offline classification accuracy.As shown in Fig. 5(a) and (c), with increasing stimulation frequency, classification accuracy and ITR rapidly improve and saturate in the low frequency band.The accuracy of ON 50% and OFF 50% drops rapidly after 20 Hz, stabilizing at 40% above 35 Hz.In comparison, the accuracy of Flicker 50% decreases noticeably after 35 Hz, while the accuracy of Flicker 300% slows down after 40 Hz and remains higher than the other three stimulation paradigms.Fig. 5(b) and (d) are pairwise comparison results of accuracy and ITR between different stimulation paradigms.Throughout the entire stimulation frequency range, there were no significant differences between ON 50% and OFF 50%.In addition, in the frequency range below 23 Hz, ON 50%, OFF 50%, and Flicker 300% all maintain comparable classification performance with no significant differences between them.After 23 Hz, Flicker 300% exhibits the highest classification performance among all four stimulation paradigms, followed by the Flicker 50% stimulation pattern.

C. User Experience Comparison Between Stimulus Paradigms
Questionnaire score results of the four stimulation paradigms at all stimulation frequencies are shown in Fig. 6.Throughout the entire frequency band, the grid stimulation has the best user experience, with Flicker 50% ranking second.As depicted in Fig. 6(b), there is no notable distinction between ON 50% and OFF 50% except at 9 Hz.And the ON grid stimulation is more comfortable than uniformly flicker stimulation at an equivalent contrast in the frequency band of 16-40 Hz.Additionally, at most stimulation frequencies, the score of Flicker 300% is significantly lower than the other three stimulation paradigms.

D. Systematic Evaluation
To systematically assess the performance and user experience of the four stimulation paradigms, different weights were used to calculate normalized accuracy and questionnaire scores, and the results are shown in Fig. 7.The advantageous stimulation paradigms vary in each frequency band when using different weights.When the ratio is 3:7, the overall score of grid stimulation is the highest in almost the whole frequency band, and with the increase of the proportion of classification performance, the frequency range in which grid stimulation dominates gradually decreases.Ensuring practicality is critical when designing stimulation paradigms for BCI systems, and it may be more reasonable to use a 7:3 ratio to select a favourable stimulation paradigm in each frequency band.That is, grid stimuli, Flicker 50%, and Flicker 300% are the optimal paradigm choices for the frequency bands of 1-25 Hz, 25-35 Hz, and 35-60 Hz, respectively.

E. BCI Performance
Based on the frequency-overall score curve in Fig. 7(c), the ten dominant stimuli (2, 5, 10, 20 Hz: ON 50% and OFF 50%; 30 Hz: Flicker 50%; 60 Hz: Flicker 300%) for different frequency bands were selected to build four-target SSVEP-based BCI systems.The online classification accuracy using the filter bank e-TRCA algorithm is shown in Fig. 8. Except for 60 Hz (79.5%, 33.26 bits/min), all classification accuracy reached over 85%, with ITR above 40 bits/min.As the stimulation frequency increases, the classification performance demonstrates an initial rise followed by a decline, with the maximum classification performance occurring at 10 Hz (ON 50%: 98.67%, 57.01 bits/min; OFF 50%: 99.67%, 59.17 bits/min).There is no significant difference between ON 50% stimulation and OFF 50% stimulation pattern under the same stimulation frequency.These results illustrate the feasibility of using the suggested dominant stimulation paradigms to build a high-performance and comfortable SSVEP-based BCI system.

F. Subjective Evaluation of BCI
The subjective evaluation results of the ten stimulation interfaces by the subjects are shown in Fig. 9(a).The user experience of the Flicker 300% stimulation pattern for 60 Hz was the best, followed by the grid stimulation mode at a frequency of 2 Hz, while the Flicker 50% with a frequency of 30 Hz had the worst experience.Fig. 9(b) shows the average ranking results of the ten stimulation interfaces across all subjects, with both accuracy and user experience taken into consideration during the ranking.Scores from 10 to 1 were assigned for the order 1-10.A higher score indicates a better stimulus interface.The user interface with Flicker 50% stimulation at 30 Hz ranks worst, which could be attributed to its worst user experience.The correlation coefficient between the simulated and actual results can be calculated when using different weights to fit the real ranking results.The correlation coefficient is the highest (R: 0.886) when using a 7:3 ratio among the three ratios as shown in Fig. 9(c).This ratio can serve as a reference for balancing performance and comfort.

IV. DISCUSSION
This study aims to optimize the combination of different stimulation frequencies and stimulation paradigms to improve the performance and user experience of SSVEP-based BCI systems.According to the comprehensive results of the frequency scanning experiment, we selected the primary stimulation paradigm for each frequency band to construct a four-target BCI system.Subsequently, the system reliability was validated by BCI experiments.
Through the frequency scanning experiment on four stimulation paradigms (i.e., ON 50%, OFF 50%, Flicker 50%, Flicker 300%) in the stimulation frequency range of 1-60 Hz, Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.the frequency-response curves (see Fig. 4) and frequency-user experience curves (see Fig. 6) of the four stimulation paradigms were obtained.The grid stimulus and uniform flicker stimulus dominated in the low-frequency band and high-frequency band, respectively.In the low-frequency range below 8 Hz, the response of grid stimuli is significantly higher than that of uniform flicker stimuli (see Fig. 3).This may be because the spatial contrast information [33] of grid stimulation induces a stronger response in the visual cortex.Previous research shows that cells in the retina and thalamic lateral geniculate nucleus have concentric ON and OFF antagonistic receptive fields, and the information processing of this characteristic receptive field emphasizes contrast information [24], [33].In electrophysiological studies, when a uniform surface stimulus of a certain luminance value is presented, some neurons do not respond to the uniform luminance stimulus [34], [35], and the response of neurons to uniform luminance is significantly weaker than the response to contrast [33].At low temporal frequencies, a uniform flicker stimulus can be approximated as a uniform luminance stimulus, while grid stimuli have more spatial contrast information than uniform flicker stimuli due to the presence of a constant luminance background region and a sinusoidally varying luminance flickering region.This may be the reason why the two peaks of grid stimuli appear earlier than those of uniform flicker stimuli in the low frequency band in Fig. 3(a).In addition, in a study investigating how cats respond to swift alterations, both in luminance and spatial contrast, nearly all V1 cells exhibited pronounced, monotonous increases in luminance tuning.This effect was particularly evident under conditions of high spatial contrast [36].The luminance flickering range of grid stimuli is significantly smaller than that of uniform flicker stimuli in this study.Perhaps as the stimulus frequency increases, the influence of stimulus luminance on evoked intensity gradually becomes dominant, and the lower flicker luminance causes the rapid attenuation of the grid stimulus response in the high frequency band.The black-background and whitebackground checkerboard-like stimuli in [10] elicited strong SSVEP responses at 40 Hz with similar spatial frequencies, but the range of flickering luminance was between the brightest white and the darkest black.It is reasonable to hypothesize that grid stimuli require higher luminance to achieve stronger responses in the high frequency band.
After the systematic evaluation of the performance and user experience of four stimulation patterns in the range of 1-60 Hz, the dominant stimulation paradigm of each frequency band was preliminarily determined (i.e., 1-25 Hz: ON 50% and OFF 50%; 25-35 Hz: Flicker 50%; 35-60 Hz: Flicker 300%).Then, the stimulation frequencies of 2, 5, 10, 20, 30, and 60 Hz were selected for the four-target SSVEP-based BCI systems, which verified the feasibility of the corresponding dominant stimulation paradigm of each frequency band.The results showed that all online classification accuracy reached more than 75% with a data length of 1 s.The sorting results of the stimulation interface showed that, except for 30 Hz, the average sorting scores of the remaining stimulation paradigms were at a comparable level.The difference in user experience scoring for 30 Hz in the BCI experiment compared to the frequency scanning experiment is likely a result of the limited variety of online stimulation paradigms and the small differences between paradigms.Additionally, the subjective assessment results can be influenced by various factors such as participants' emotions, age, gender, and others.To acknowledge the potential impact of these variables on the results, a more comprehensive investigation of these factors could be explored in future research.Table I summarizes the stimulation paradigms, frequencies, and performance of userfriendly four-target BCI systems in recent years.Research methods to improve user experience in low and medium frequency bands include applying spatial encoding strategies to reduce the number of stimuli [37] or designing new stimulus paradigms [20], [21], etc.This study achieved higher classification accuracy and ITR than previous studies by using grid stimulation paradigms with high comfort in low and medium

TABLE I DIFFERENT USER-FRIENDLY FOUR-TARGET BCI SYSTEMS
frequency bands.In the high frequency band, most studies still adopt the traditional black-and-white flickering stimulation paradigm in 30-40 Hz due to the weak flicker perception in this frequency range [38], [39], [40], [41].This study achieved improvement in performance and user experience by implementing the Flicker 50% stimulation pattern at 30 Hz, which provided a more comfortable experience compared to traditional uniform stimuli.At a stimulation frequency of 60 Hz, most existing studies have employed the traditional black-and-white stimulation paradigm [16], [32].Ming et al. implemented an individualized space and phase modulation strategy [32] to achieve an ITR of up to 52.8 bits/min, which can be applied in future research.
In future studies, the effect of other stimulation attributes of grid stimulation, such as stimulation regional proportion, contrast, spatial frequency, etc., on the SSVEP characteristics will be further investigated.Meng et al. reported that stimuli with 20%-pixel density can help balance performance and user experience [42].Reducing the proportion of flickering areas in the entire pattern can decrease the overall perception of flicker, thereby contributing to further enhancing the user experience in the low and middle frequency bands.Additionally, contrast and spatial frequency attributes have previously been shown to have a significant impact on response in similar studies [10], [21], and further optimization of these parameters may improve classification performance in the high frequency band.Given the technological advancements in LCD screen refresh rates reaching 360 Hz [16] and the development of more diversified presentation technologies, such as augmented reality [43], the proposed stimulation paradigms can further enrich the application scenarios of BCI systems.

V. CONCLUSION
This study systemically explored the influence of stimulation frequency on the SSVEP characteristics and user experience of four stimulation paradigms (i.e., ON 50%, OFF 50%, Flicker 50%, Flicker 300%).By using a weighting ratio of 7:3 to allocate the simulated classification accuracy and subjective score, the four stimulation paradigms were systematically evaluated, and the optimal stimulation pattern Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
was selected for each frequency band.The effectiveness of selecting stimulation paradigms through this strategy has been validated to a certain extent in the four-target online SSVEPbased BCI system.Grid stimulation can replace traditional uniform flicker stimulation in low and medium frequency bands to attain a better trade-off between performance and user experience.In the high frequency range below 35 Hz, uniform flicker stimulation with low contrast is the best choice.These findings advance the development of more effective and userfriendly SSVEP-based BCIs for practical applications.

Fig. 1 .
Fig. 1.The grid and uniformly flickering stimulation paradigms with high and low contrasts used in the experiment.The y-axis represents the luminance value used in the experiment, labeled in unequal proportions for better presentation.

Fig. 3 .
Fig. 3. Average SSVEP signals across all subjects at 2, 12, 30, and 45 Hz for the four stimulus patterns.(a) SSVEP signals at the Oz electrode with a duration of 1.0 s.(b) The topographic maps on the amplitude of SSVEPs at the fundamental frequency.(c) The mean amplitude spectra of SSVEPs across nine electrodes.(d) The mean SNR spectra of SSVEPs across nine electrodes.The scale for each subplot is determined by taking the maximum value among the four paradigms.All data epochs were filtered using a [f-0.2100] Hz bandpass filter, where f represents the four stimulus frequencies.

Fig. 4 .
Fig. 4. The amplitude and SNR of SSVEPs for the four stimulus patterns.(a) The fundamental amplitude varies with stimulation frequency.(b) The SNR at the fundamental frequency varies with stimulation frequency.(c) The amplitude at the second harmonic varies with the stimulation frequency.(d) The SNR at the second harmonic varies with the stimulation frequency.The shaded area represents the standard error.

Fig. 5 .
Fig. 5. (a) Offline classification accuracy and (c) ITR of the four stimulation paradigms at all stimulation frequencies.The shaded area represents the standard error.(b) and (d) are pairwise comparison results of accuracy and ITR between stimulation paradigms, respectively.The colour bar represents the p-value, and * represents p<0.05.

Fig. 6 .
Fig. 6.The mean scores of the questionnaire evaluation for the three criteria (i.e., comfort level, flicker perception, and preference) across the four stimulation paradigms, as well as pairwise comparison differences between stimulation paradigms.The shaded area in (a) represents the standard error.The colour bar in (b) represents the p-value, and * represents p<0.05.

Fig. 8 .
Fig. 8. (a) Online classification accuracy and (c) ITR of the ten distinct SSVEP-based BCI systems.(b) and (d) are pairwise comparison results of accuracy and ITR between stimulation paradigms, respectively.The colour bar represents the p-value, and * represents p<0.05.

Fig. 9 .
Fig. 9. (a) The average user experience scores for the ten stimulation interfaces.(b) The average ranking results of the ten stimulation interfaces across all subjects.The blue solid line represents the real ranking result and the other dashed lines represent simulated ranking results using different weighting ratios (i.e., 1:1, 3:7, and 7:3) between online performance and user experience.(c) The correlation coefficient between the simulated result and the true result using different weights.