FIG. 39: Single channel correlation activations for each of the 25-target positions averaged across all subjects. T he correlations are plo tted topographically on head- models th a t are positioned a t each of th e 25-target positions. Each head-m odel is scaled to the m inim um and m axim um correlation value of the respective targ et position. The dark-red color indicates areas of high contributing activity for the c- V E P responses. It can be seen th a t the responses from direct foveal stim ulation (outer ring) exhibit centrally over the prim ary visual cortex while parafoveal responses are exhibited from the parietal-occipital regions of th e visual cortex.
7.4 DISCUSSIO N
T he m ain objective of this offline analysis was to test th e feasibility of 25-target detection using only 4 m-sequence stimuli. Currently, alm ost all V EP-based BCIs utilize a separate stimulus for each targ et or class. This work has th e potential for not only leading to more practical V EP-B CIs from th e reduction visual fatigue and irritatio n by reducing the num ber of flashing stimuli, b u t it can also potentially lead to th e development of BCIs w ith higher ITRs. The average IT R of 88 b its /m for a single cycle observation length is com parable to th e IT R results obtained in G ao’s cornerstone c-V EP paradigm [Bin et al., 2011) b u t w ith only a fraction of the num ber of flashing stim uli used. T he best subject was able to achieve an IT R of over 120 b its/m in , which outperform s G ao’s results. However, the present results need to be confirmed w ith an online im plem entation.
Nevertheless, these results represent an im provement in practicality as th e a sig nificant reduction of flashing stimuli is achieved. The target-to-stim ulus ratio for the current ring paradigm is 6.25 targets per stimulus (25 targ ets / 4 stimuli) whereas th e G ao’s paradigm has a ratio of 0.53 targets per stimuli (32 targ ets / 60 stimuli). To th e a u th o r’s b est knowledge, no study to d ate has attem p ted to increase the num ber of available targets to be larger th a n the num ber of flashing stimuli. Therefore, although th e current results are from offline analysis and need to be repeated online, th e obtained classification perform ance shows promise th a t V EP BCIs can utilize spatial decoupling of the targets and flashing stimuli to gain 3-fold increase in the num ber of targ et classes w ithout increasing th e num ber of flashing stimuli.
T he concept of spatial decoupling can potentially increase the num ber of targ ets p ast a given physical limit on the bottlenecks th e num ber of flashing stim uli for other existing paradigm s. For example, G ao’s paradigm owes its high inform ation transfer rate to the 32 targ ets (and 32 + 28 corresponding flashing stimuli) it employs. However, th a t paradigm is unable to increase the num ber of targets further as the length of th e m-sequence is unable to employ more th a n 32 distinct sim ultaneous flashing stimuli. Using this spatial decoupling technique, it is conceivable th a t G ao’s
paradigm could be extended to employ more th a n 32 targ ets w ithout increasing the num ber of flashing stim uli which would potentially lead to higher inform ation transfer rates.
In addition to the 25-target detection, the present characterization analysis of th e c-VEP responses elicited from th e ring paradigm helps gain further insight on th e n atu re of spatial decoupling. For instance, Figure 38 shows th a t when visually attending to (or near) the boundary of two adjacent m-sequence stimuli, the result ing c-VEP response is a com bination of roughly equal proportions of th e individual responses elicited innately by those two stimuli. Further, a predictable decrease in the response stren g th is shown as the distance between th e targ et and stim ulus is increased. A m athem atical model representing V EP spatial decoupling can be made by modeling the correlation stren g th as a function of target-stim ulus distance and as a function targ et adjacency relationships. This model could potentially be used as m ethod to reduce the am ount of training tim e needed for the BCI system. Currently, training d a ta is needed from each targ et position in the BCI for construction of the tem plate responses. A model describing the stim ulus-target spatial decoupling could allow for a single m aster tem plate to be made, requiring training d a ta from only one target, and then constructing the rem aining targets from th e m aster tem plate. Future analysis will test th e efficacy of such model construction for BCI purposes.
Similar to the source correlations, the head-model topographies of the c-VEP responses shown in Figure 39 dem onstrate the spatial activations th a t are elicited from th e asym m etrical stim ulation of th e ring stim ulus paradigm . The topographies show a noticeable p a tte rn where targ ets th a t are positioned directly over th e ring stimulus result in strong medial activations of th e occipital lobe and targets th a t are positioned further away in th e parafoveal stim ulation region show more lateral activations. On th e horizontal plane, th e left and right parafoveal targ ets show strong right and left lateralizations, respectively. In future analysis, the knowledge channel- specific activations due to target-stim ulus positional relationships can be used as a priori inform ation for multi-class targ et detection.