Optimizing SSVEP-based BCI training through Adversarial Generative Neural Networks

Brain-computer interfaces (BCIs) based on steady-state visually evoked potential (SSVEP) use brain activity to control external devices, with applications ranging from assistive technologies to gaming. Typically, BCI systems are developed using supervised learning techniques that require labelled brain signals. However, acquiring these labelled signals can be tiring and time-consuming, especially for subjects with disabilities. In this study, we evaluated the performance impact of using synthetic brain signals to train and calibrate an SSVEP-based BCI system. Specifically, we used generative adversarial networks (GANs) to synthesize brain signals with SSVEP information, considering cases with two and four visual stimuli. Four scenarios with different proportions of real vs. synthetic brain signals were evaluated: Scenario 1 (baseline) using only real data and Scenarios 2-4 with 10%, 20% and 30% of real data replaced by synthetic data, respectively. Our results reveal that synthetic data can be used to train the BCI without a performance loss across the tested scenarios when two visual stimuli are used and with an average performance reduction compared to baseline of 7% (Scenario 2), 10,3% (Scenario 3) and 9,3% (Scenario 4) for four stimuli. Furthermore, considering each recording has duration of 2 seconds, by replacing 30% of real data with synthetic data, there is an immediate time-saving of 48 s and 96 s in the cases with two and four visual stimuli, respectively. This trade-off between accuracy and efficiency has significant implications for improving the usability and accessibility of SSVEP-based BCI, especially for assistive applications.