Gaze-Independent Task Performance

There is only one published covert attention study based on c-VEP that we are aware of. In addition to the overt task mentioned above, (Waytowich and Krusienski, 2015) looked at near-foveal (1° visual angle away) and parafoveal (4° visual angle away) target stimuli. Instead of having users fixate on a central point and attend to one of multiple equidistant targets, however, this study has the users attend to fixation points close to but not on top of a region of a large annulus-based stimulus. They were able to achieve group average accuracy above 80% for a trials based on 6 cycles (6.3s) and at

least 70% for some individuals on only single-cycle classification attempts. While this indicates that c-VEP is generally detectable with covert attention, it does so in a paradigm that is still actually gaze-dependent. Furthermore, it is unclear how well this approach would stand up in even a binary discrimination task if both stimuli were within the same visual angle at the same time.

Covert visual attention BCI studies based on an SSVEP paradigm have been reported as well (Kelly et al., 2005; Walter et al., 2012). The effects of covert attention were detectable and had contralateral representations in left-right binary tasks. However, the power of the detected signals was much smaller and resulted in around a 20% drop in classification accuracies compared to overtly-attended versions of the same task.

The covert c-VEP methods attempted here failed to produce any usable level accuracies and only a few instances barely managed to show significance above chance with one exception from the adaptive online 2DLLAP scenario, though that was still well below usability threshold. One challenge was the placement of the stimuli themselves. In an attempt to align the stimuli with the BCI application workspace, they were placed farther out (5.7° to inner edge) than has been reported for covert VEP studies, which range from 1° to 4.9°. This is, however, well within the range of covert vision-based event-related potential (ERP) BCI studies (Treder et al., 2011; Marchetti et al., 2013; Martel et al., 2014), which placed the center of the stimuli up to 10° from center fixation. Moving the stimuli closer to center may help improve the SNR, though would also likely require a redesign of how the application workspace was laid out. From a different perspective, it may have been that the stimuli were actually too close together. With all

four stimuli visible and occupying roughly the same amount of visual space, only top- down attentional effects were likely introducing differences in the trials from the combined flicking activity from all stimuli. Trying to remove this baseline through subtractive means proved to actually decrease performance, as the subtraction of a common vector from each template only increased their correlations with each other, thus making correlation-based discrimination even more difficult.

Another challenge was the limited duration of the trials used in building and evaluating the classifiers, as only single-cycle trials of 1.034s were evaluated. The experiment was designed to enable up to three-cycle trials in online mode, so further analysis needs to be done to see if using two- or three-cycle trials can improve performance.

The overlapped c-VEP task showed some promise, as three subjects were able to perform greater than the usability threshold of 70% accuracy in offline analysis while two of them maintained that level online as well. While the average group accuracy was not high, the fact that some subjects were able to successfully perform the task at a usable level suggests that this method bears further investigation.

A major challenge to this approach is that, due to the fact that both stimuli were present at the same time regardless of gaze, templates were highly correlated, even more so than those generated for the covert task. Some form of decorrelation method needs to be applied prior to classification, or a non-correlative classifier employed to overcome this issue. Likewise, only singe-cycle trials were evaluated, so extending the analysis to two- or three-cycle trials may prove fruitful as well.

Many covert visual attention BCI studies refer to the work as an independent BCI because gaze control is not required. However, these studies assume that, while reliable gaze movement is unavailable, gaze will remain fixed at a set location. Some motor impaired patients may have involuntary gaze movement or may not be able to maintain gaze directed squarely at a central fixation point. Some work has gone into non-spatial visual selective attention where the whole screen is used to present two overlapping SSVEP stimuli. Zhang et al. demonstrate a method where gaze is fixed at a central spot while attention is directed to one of two overlapping stimuli: red and blue dots that flicker at different rates and rotate around the fixation point counterclockwise and clockwise, respectively (Zhang et al. 2010). The rotation allows for the perception of two separate planes, similar to the checkerboard offset and transparency method used here. They were able to achieve an average accuracy of 72% across 18 subjects. Allison et al. also explored overlapping stimuli for SSVEP, with alternating red and green lines. They did not constrain eye gaze, but were also unable to find high performance accuracy.

This work is the first we are aware of to explore not only a gaze-independent c- VEP paradigm, but also potentially the first of any reported visual BCI paradigm that allows the gaze to be anywhere on the task space. Having the entire display taken up by flickering checkerboards does present a challenge for practical application use, however. The intent is that, by using slightly transparent stimuli, application-specific information can be displayed underneath the stimuli and so still be visible.