Objective: Steady-state visual evoked potential (SSVEP) is a very popular approach to establishing a communication pathway in brain-computer interfaces (BCIs), without any training requirements for the user. Brain activity recorded over occipital regions, in association with stimuli flickering at distinct frequencies, is used to predict the gaze direction. High performance is achieved when the analysis of multichannel signal is guided by the driving signals. This study introduces an efficient way of identifying the attended stimulus without the need to register the driving signals.
Approach: Regional brain response is described as a dynamical trajectory towards one of the 'attractors' associated with the brainwave entrainment induced by the attended stimulus. A condensed description for each single-trial response is provided by means of discriminative vector quantization, and different trajectories are disentangled based on a simple classification scheme that uses templates and confidence intervals derived from a small training dataset.
Main results: Experiments, based on two different datasets, provided evidence that the introduced approach compares favorably to well-established alternatives, regarding the information transfer rate.
Significance: Our approach relies on (but not restricted to) single sensor traces, incorporates a novel description of brainwaves based on semi-supervised learning, and its great advantage stems from its potential for self-paced BCI.