A Brain Switch for SSVEP-Based BCI Speller Using an RNN-Based Detection Approach

Steady-state visual evoked potentials (SSVEP)-based brain-computer interface (BCI) systems are used commonly as spellers because they have high information transfer rate and high accuracy relative to other BCI paradigms. Asynchronous BCI systems allow users to input commands whenever they wish to use them, which may make these systems more realistic and practical than synchronous systems. In contrast, asynchronous BCIs, known as the Brain Switch, require robust mechanisms to detect users' intentions accurately while maintaining classification performance. This highlights the need for a BCI system that distinguishes users' intentions reliably. SSVEP paradigms often show variability in their frequency designs. In this study, we propose a two-stage asynchronous BCI system that combines a robust brain switch model that uses autocorrelation and Long Short-Term Memory (LSTM)) for detection and an EEGNet-based classifier. Our proposed system was evaluated using a 40-class SSVEP dataset involving 40 subjects. It achieved an impressive detection performance with a sensitivity (SEN) of 98.24 ± 2.21% and specificity (SPC) of 82.28 ± 11.63% for even 1-second epochs. Further, the system attained a classification accuracy (ACC) of 77.05 ± 14.95%. This model demonstrates significant potential to help develop more realistic and practical asynchronous BCI systems.