110 120 F 1 30 140 150 160 170 0 0 .1 5 0 .3 0 .6 1.2 2 .4 4 .8 9 .6 19 .2
Spatial Frequency (Cycles/Deg)
(a)
Path Completion Percentage
100 9 5 CL 8 5 8 0 7 5 7 0 0 0 .1 5 0 .3 0 .6 1.2 2 .4 4 .8 9 .6 1 9 .2Spatial Frequency (Cycles/Deg)
(b)
FIG. 29: P a rt (a) shows the average run-duration from experim ent 2 across all sub jects for each of th e spatial frequency conditions. P a rt (b) shows th e average path- com pletion percentage for each of th e conditions. On average, th e subjects were able to complete th e p a th th e fastest w ith the 0 c/d eg and 2.4 c/d eg conditions which is reflected in (b) as those were th e only two conditions were all subjects were able to completely finish the path.
5.4 DISCUSSIO N
As the effect of spatial frequency on BCI perform ance has not been explored, m ost studies arbitrarily select stim uli using either a solid image or a generic checker board p a tte rn in th e range of 0.15-0.3 c/deg. This study dem onstrates th a t spatial frequency can have a dram atic effect on SSVEP perform ance th a t is consistent across subjects. The results in Figure 24 show th a t the solid 0 c/d eg condition is able to achieve an average accuracy of 97.7% given observation lengths greater th a n 3.5 s. The 2.4 c/d eg spatial condition is also able to achieve a reasonable accuracy of 85.1% using a shorter observation length of 2.5 s. Even though th e 2.4 c/d eg accuracy is not a high as th a t obtained from the solid condition, the responses generated from the 2.4 c/d eg spatial frequency condition require a shorter tim e window for excitation. This is reflected in Figure 25 where the 2.4 c/deg condition obtains an averaged IT R of 45.3 b its/m in which is significantly higher com pared to 35.7 b its/m in w ith th e solid stim ulus condition (p=0.02).
The results from b o th the discrete classification and th e continuous navigation experim ents show a clear bim odal distribution of SSVEP perform ance across the spatial frequency conditions. Figures 24, 25, and 26 from the discrete classification experim ent all reflect a similar bim odal spatial tuning where perform ance peaks are exhibited around the 0 c/d eg and 2.4 c/d eg conditions. For th e accuracy plots, the peaks a t the 0 c/d eg condition are higher th an th e peaks a t th e 2.4 c/deg, whereas th e opposite is seen from th e IT R plots where the 2.4 c /d eg peaks exceed the 0 c/d eg peaks. The continuous navigation experim ent confirms th e results from experim ent 1 as Figure 29 shows th a t the perform ance of the navigation task averaged across all subjects exhibits a similar bim odal distribution across spatial frequency condi tions. The 0 c/d eg and 2.4 c/d eg peaks were the only conditions where all subjects were able to fully complete th e path, thus giving a strong indication th a t these two spatial frequency conditions provide superior response characteristics and can result in optim al SSVEP perform ance com pared to other spatial frequencies. These re sults concur w ith the findings in [Tobimatsu et ah, 1993, Tornoda et ah, 1991] where
similar bim odal spatial tuning curves were found when m easuring V E P am plitude during stim ulation from a single stimulus. This study shows for th e first tim e th a t a similar bim odal relationship occurs between SSVEP BCI perform ance and spatial frequency even when multiple stim uli are flashed sim ultaneously a t different tem poral frequencies.
Interestingly, th e 0 c/deg condition resulted in a faster com pletion tim e on aver age, com pared to the 2.4 c/d eg condition despite the fact th a t th e 2.4 c/d eg achieves an overall higher IT R th a n 0 c /d eg condition. This m ay be due to the decrease in accuracy exhibited by the 2.4 c/d eg condition for longer observation lengths, as seen in Figure 27. Additionally, all of th e spatial frequency conditions, except for th e solid condition, show a substantial decrease in accuracy for observation lengths longer th a n 3 seconds th a t continues to decrease as the observation length increases. This suggests th a t a mechanism of spatial ad ap tatio n m ight be occurring in the vi sual system from the stim ulation of these spatial frequency conditions during th e SSVEP BCI tasks. It is a known phenomenon th a t there exists a reduction of neural activity when stimuli are continuously repeated [Grill-Spector et al., 2006]; however, the underlying neural mechanisms of this phenomenon are still unknown. In th e case of spatial frequency ad aptation, there exists spatially tu n ed neuronal popula tions (or spatial channels) th a t can ad ap t during stim ulation of spatial stimuli in which the strength of the spatial channel response declines throughout ad ap tatio n [Blakemore and Sutton, 1969, Klein et ah, 1974, Movshon and Lennie, 1979]. The results from this current study show a similar p a tte rn of spatial frequency a d a p ta tion over th e time-course of stim ulation th a t agrees w ith previous EEG studies th a t use a single stim ulus [Heinrich and Bach, 2001, Baas et al., 2002], T he results from this current study show th a t the mechanisms of spatial a d ap tatio n similarly occur in the context of SSVEP BCIs where m ultiple sim ultaneously flashing stim uli are presented to th e visual system. These have im portant im plications for V EP-based BCIs where sim ultaneously flashing stimuli are common. Thus, when using spatial frequency stimuli, the targ et detection accuracy over prolonged durations of stim u lation can decrease due to spatial ad ap tatio n - a fact th a t is generally overlooked in
V EP-B CI studies.
The results in Figure 28 of th e subjective evaluation of visual irritatio n for each spatial frequency condition show th a t subjects, on average, perceive less overall ir ritatio n for higher spatial frequencies com pared to lower spatial frequencies. Specif ically, for th e two to p perform ing conditions, th e average irritatio n index was 8.1 for the 0 c/d eg condition, whereas the irritatio n index was roughly half th a t a t 4.5 for the 2.4 c /d eg condition. This has favorable implications which show th a t p racti cal V EP-based BCIs can be employed w ith less visually irritating stim uli to achieve com parable perform ance w ith th e traditional, more obtrusive solid stimuli.
Although these results dem onstrate th a t spatial frequency exhibits a distinct p a t tern on th e accuracy and perform ance of SSVEP BCI targ et detection, additional analysis is still needed to further characterize these effects. A longitudinal study is needed to determ ine the stability of subject-specific spatial frequency tuning and adaptation. Further, the relationship between spatial frequency and th e tem poral flashing (or p a tte rn reversal) frequency needs to be studied as th e effect of tem poral frequency on spatial frequency tuning and adap tatio n is not well-understood. Overall, these results dem onstrate th a t th e clinically studied mechanisms of spatial frequency tuning and ad ap tatio n are present in the context of m ulti-target stim ula tion, showing th a t spatial frequency plays a significant role in SSVEP performance. This characterization can potentially be utilized for th e developm ent of more p racti cal and robust BCIs.