Identifying potential training factors in a vibrotactile P300-BCI

Abstract

Brain-computer interfaces (BCI) often rely on visual stimulation and feedback. Potential end-users with impaired vision, however, cannot use these BCIs efficiently and require a non-visual alternative. Both auditory and tactile paradigms have been developed but are often not sufficiently fast or accurate. Thus, it is particularly relevant to investigate if and how users can train and improve performance. We report data from 29 healthy participants who trained with a 4-choice tactile P300-BCI during five sessions. To identify potential training factors, we pre-post assessed the robustness of the BCI performance against increased workload in a dual task condition and determined the participants' somatosensory sensitivity thresholds with a forced-choice intensity discrimination task. Accuracy (M = 79.2% to 92.0%) and tactually evoked P300 amplitudes increased significantly, confirming successful training. Pre-post somatosensory sensitivity increased, and workload decreased significantly, but results of the dual task condition remained inconclusive. The present study confirmed the previously reported feasibility and trainability of our tactile BCI paradigm within a multi-session design. Importantly, we provide first evidence of improvement in the somatosensory system as a potential mediator for the observed training effects.

Figure 1

A C2 tactor (by Engineering Acoustic Inc., Casselberry, USA) and the four stimulus positions.

Figure 2

Intensity discrimination example, indicating the ratio of equal/unequal responses from one participant. Here, responses were ambiguous in the range between 40% and 25% intensity differences. We defined the intensity at which both answers would be given with the same probability (0.5) as the discrimination threshold. This point was estimated using a sigmoidal fit.

Figure 3

Grand average ERP data. Top: ERP shape comparison between first and last session at Fz, Cz and Pz (calibration and copy task). Bottom: Extracted mean values (amplitude & curve difference) across all sessions. Shaded areas and error bars indicate standard error of the mean (SEM).

Figure 4

Development of BCI accuracy and comparison of copy and dual task conditions across sessions. Left: Average BCI accuracies achieved during the respective tasks. Shaded areas and error bars indicate SEM. Right: Target vs. non-target epochs at Fz from copy and dual tasks.

Figure 5

Analysis of the somatosensory sensitivity. After training, participants were able to discriminate between smaller stimulus differences, indicated by significantly lower sensory thresholds. There was a negative correlation between the sensory thresholds and accuracy in session one, but not session five.