Training the spatially-coded SSVEP BCI on the fly

BACKGROUND: The spatially-coded SSVEP BCI employs the retinotopic map in the human visual pathway to infer the gaze direction of the operator relative to a flicker stimulus inducing steady-state visual evoked potentials (SSVEPs) in the brain. It has been shown that with this method, up to 16 channels can be encoded using only a single flicker stimulus. Another advantage over conventional frequency-coded SSVEP BCIs, in which channels are encoded by different combinations of frequencies and phases, is that the operator does not have to gaze directly at flickering lights. This can reduce visual fatigue and improve user comfort. Whereas the frequency of the SSVEP response is well predictable, which has enabled the development of frequency-coded SSVEP BCIs which do not require training data, the spatial distribution of the SSVEP response over the scalp differs much more between different people. This requires collecting a substantial amount of training data before the spatially-coded BCI could be put into operation.

NEW METHOD: In this study we address this issue by combining the spatially-coded BCI with a feedback channel which the operator uses to flag classification errors, and which allows the system to accumulate valid training data while the BCI is used to solve a spatial navigation task.

RESULTS: Starting from the minimal number of samples required by the classification method, the approach achieved an average accuracy of 69 ± 15 %, corresponding to an ITR of 31 ± 17 bits/min, in solving the task for the first time. This accuracy improved to 87 ± 9 % (ITR: 54 ± 14 bits/min) after completing the task 2 more times. Further we show that participants with a stable SSVEP topography over repeated stimulation enable the BCI to achieve higher accuracies.

COMPARISON WITH EXISTING METHODS: Compared to a similar system with separate training and application phases, the time to achieve the same output is reduced by more than 50 %.

CONCLUSIONS: Evaluating the approach in 17 participants suggests that the performance of the spatially-coded BCI with a minimal set of training samples is sufficient to be operational, and that performance keeps improving in the course of its application.