Publication:

A fast SSVEP-based brain-computer interface

Literature of brain-computer interfacing (BCI) for steady-state visual evoked potentials (SSVEP) shows that canonical correlation analysis (CCA) is the most used method to extract features. However, it is known that CCA tends to rapidly overfit, leading to a decrease in performance. Furthermore, CCA uses information of just one class, thus neglecting possible overlaps between different classes. In this paper we propose a new pipeline for SSVEP-based BCIs, called corrLDA, that calculates correlation values between SSVEP signals and sine-cosine reference templates. These features are then reduced with a supervised method called shrinkage linear discriminant analysis that, unlike CCA, can deal with shorter time windows and includes between-class information. To compare these two techniques, we analysed an open access SSVEP dataset from 24 subjects where four stimuli were used in offline and online tasks. The online task was performed both in control condition and under different perturbations: listening, speaking and thinking. Results showed that corrLDA pipeline outperforms CCA in short trial lengths, as well as in the four additional noisy conditions.