Effects of Training on BCI Accuracy in SSMVEP-based BCI

Abstract

This paper investigates the effects of the training process on the classification accuracy for a steady-state motion visual evoked potentials (SSMVEP)-based brain-computer interface (BCI) paradigm.

An SSMVEP-based BCI works similar to SSVEP with the main difference that the stimulus is smoothly changing its appearance, with a continuous motion, leading to less user fatigue. Typical SSMVEP classification utilises correlation algorithms to compare the incoming Electroencephalography (EEG) data with a sine-cosine template.

To increase the classification performance of BCI algorithms, collecting user training data has been a common practice recently, usually for template-matching detection algorithms. The incoming EEG data are compared with an individually created template from the user’s own pre-recorded EEG response to the stimulus. In this offline study, previously recorded data (3 s of training EEG data), which were collected during an online experiment with 86 participants, were used.

Task-related component analysis (TRCA), a state of the art classification method, was modified with the spatial filter W generated by the canonical-correlation analysis (CCA). The TRCA and the sine+cosine templates were compared. The cross paradigm utilisation of the training data was also investigated, e.g. the TRCA model built from SSVEP training data was used to classify the SSMVEP data and vice versa.

Results show a significant difference in favour of the usage of the training data over the sine-cosine template for the SSMVEP paradigm classification. A cross-paradigm validation shows promising results (accuracy >70%) for a time window of 1.5 s, similar to the sine+cosine templates. The trained TRCA models achieved an accuracy of 94% and 97% for the SSMVEP and SSVEP paradigms, respectively, in the mentioned time window. Overall, we conclude that training can significantly improve the SSMVEP target classification.