A BCI Gaze Sensing Method Using Low Jitter Code Modulated VEP

Abstract

Visual evoked potentials (VEPs) are used in clinical applications in ophthalmology, neurology, and extensively in brain-computer interface (BCI) research. Many BCI implementations utilize steady-state VEP (SSVEP) and/or code modulated VEP (c-VEP) as inputs, in tandem with sophisticated methods to improve information transfer rates (ITR). There is a gap in knowledge regarding the adaptation dynamics and physiological generation mechanisms of the VEP response, and the relation of these factors with BCI performance. A simple, dual pattern display setup was used to evoke VEPs and to test signatures elicited by non-isochronic, non-singular, low jitter stimuli at the rates of 10, 32, 50, and 70 reversals per second (rps). Non-isochronic, low-jitter stimulation elicits quasi-steady-state VEPs (QSS-VEPs) that are utilized for the simultaneous generation of transient VEP and QSS-VEP. QSS-VEP is a special case of c-VEPs, and it is assumed that it shares similar generators of the SSVEPs. Eight subjects were recorded, and the performance of the overall system was analyzed using receiver operating characteristic (ROC) curves, accuracy plots, and ITRs. In summary, QSS-VEPs performed better than transient VEPs (TR-VEP). It was found that in general, 32 rps stimulation had the highest ROC area, accuracy, and ITRs. Moreover, QSS-VEPs were found to lead to higher accuracy by template matching compared to SSVEPs at 32 rps. To investigate the reasons behind this, adaptation dynamics of transient VEPs and QSS-VEPs at all four rates were analyzed and speculated.

Keywords: BCI; QSS-VEP; SSVEP; c-VEP; deconvolution; gaze sensing; transient VEP.

Figure 1

Overall system setup and the experimental procedure are shown for a 10 rps stimulation. The stimulator is shown at the top left, driving sequences for the left field (LF) and right field (RF) are shown at the bottom left, the quasi-steady-state (QSS) and transient responses from subject 5 are shown in the bottom right figures. Interference signals after averaging 127/128 sweeps in one file for left/right, $y_{RL}\left(t\right)$ and $y_{LR}\left(t\right)$, are also displayed to highlight the orthogonality of the sequences. Refer to Section 2.4 and abbreviation list at the end of the paper for the abbreviations used in the figure.

Figure 2

Population average left field quasi-steady-state-visual evoked potentials (QSS-VEP) (A), and transient-visual evoked potentials (TR-VEP) (B) plots are shown. (A) 10 rps, 32 rps, 50 rps, and 70 rps QSS-VEPs are at the left. (B) TR-VEPs are at the right. Note the morphology change in the TR-VEPs by rate from 10 rps to 32 rps.

Figure 3

Subject 3 (S3) single sweep 0.5-second condition. (a) Left QSS-VEP receiver operating characteristic (ROC) curves and areas for all the rates are shown at the left. (b) Left TR-VEP ROC curves and areas for all the four rates are shown at the right.

Figure 4

QSS-VEP performances for left and right gazes were averaged as single QSS-VEP performance parameter at each rate. For single sweep 0.5 s, 1 s, and 1.5 s, 32 rps had the highest accuracy and ITR performance (p < 0.05) with QSS-VEP signature while 70 rps had the lowest performance (p < 0.05). For the eight subject population, average ITR values for QSS-VEP (top-right) and TR-VEPs (bottom-right) were plotted for each rate and data durations.

Figure 5

QSS-VEP and steady-state visual evoked potentials (SSVEP) performances are compared. QSS-VEPs performed better than SSVEPs for short data durations.