Abstract
We have developed an unsupervised multi-class classification algorithm of Steady-State Visual Evoked Potential (SSVEP)-based-BCI. Our previously proposed algorithm detects the frequency that gives maximum peak-to-baseline amplitude of the spectrum power among the frequency bands corresponding to the flicker light stimuli at the three electrodes arranged in occipital area, and majority voting among these electrodes determines the final output of the classifier. To further increase the accuracy, we examined whether (1) a feature that incorporates the peak-to-baseline amplitude of the spectrum power at the harmonic frequencies of the flicker frequency in addition to those at the fundamental frequency, and/or (2) a winner-take-all fashion of classification algorithm (classification based on the feature at a single electrode showing the maximum peak-to-baseline amplitude among three electrodes), could improve the accuracy of the classifier. The best accuracy (79.2%, N=32) was obtained when we incorporated the peak-to-baseline amplitude at the first and the second harmonics into features and selected the frequency at which the peak-to-baseline amplitude showed the largest value among the stimuli frequencies and electrodes. With this improved feature selection and classification algorithm, we could significantly improve the accuracy of the classifier compared to our previous algorithm (68.8%, p<0.001).
