Dynamic time window mechanism for time synchronous VEP-based BCIs-Performance evaluation with a dictionary-supported BCI speller employing SSVEP and c-VEP

Abstract

Brain-Computer Interfaces (BCIs) based on visual evoked potentials (VEPs) allow high communication speeds and accuracies. The fastest speeds can be achieved if targets are identified in a synchronous way (i.e., after a pre-set time period the system will produce a command output). The duration a target needs to be fixated on until the system classifies an output command affects the overall system performance. Hence, extracting a data window dedicated for the classification is of critical importance for VEP-based BCIs. Secondly, unintentional fixation on a target could easily lead to its selection. For the practical usability of BCI applications it is desirable to distinguish between intentional and unintentional fixations. This can be achieved by using threshold-based target identification methods. The study explores personalized dynamic classification time windows for threshold-based time synchronous VEP BCIs. The proposed techniques were tested employing the SSVEP and the c-VEP paradigm. Spelling performance was evaluated using an 8-target dictionary-supported BCI utilizing an n-gram word prediction model. The performance of twelve healthy participants was assessed with the information transfer rate (ITR) and accuracy. All participants completed sentence spelling tasks, reaching average accuracies of 94% and 96.3% for the c-VEP and the SSVEP paradigm, respectively. Average ITRs around 57 bpm were achieved for both paradigms.

Fig 1. Example of the proposed automatic selection of the classification parameters.

Displayed are the results from the leave-one-out off-line analysis of the SSVEP training data for one participant. In the training session, each of the 8 targets was attended 6 times for 3 seconds. (A) Displayed are the correlation values of the target stimulus (green) and the maximum correlation of the non-target stimuli (red). For each stimulus class the correlations calculated in the first second are displayed. (B) ITR averaged over the 6 training blocks. Here, the time window yielding the highest ITR was 0.45 seconds. (C) Correlogram of the training data. Depicted are the correlations for each target using the determined time window, averaged over the trials. This minimal difference between target and non-target stimuli was used as the threshold value. Here, the distance was minimal for the 10 and 14 Hz pair, yielding a difference of 0.2.

Fig 2. Writing a sentence with the dictionary-driven speller.

A participant is writing JUST_DO_IT in seven steps. Selection of individual letters required two steps: Initially the group containing the character was selected (Layer I), and then the desired character was selected (Layer II).

Fig 3. Comparison of SSVEP and c-VEP methods.

ITRs and accuracies of all participants are presented. The asterisks mark statistical significance determined with paired sample t-tests (*p < 0.05 and **p < 0.01).

Fig 4. Results from the user questionnaires.

Responses were given on a 1-5 Likert scale, 1 indicating strong disagreement and 5 indicating strong agreement.