Abstract
Non-stationarity in EEG signals poses significant challenges for the performance and implementation of brain–computer
interfaces (BCIs). In this study, we propose a novel method for cross-session BCI tasks that employs a supervised
autoencoder to reduce session-specific information while preserving task-related signals. Our approach compresses high-
dimensional EEG inputs and reconstructs them, thereby mitigating non-stationary variability in the data. In addition to
unsupervised minimization of the reconstruction error, the objective function of the network includes two supervised terms to
ensure that the latent representations exclude session identity information and are optimized for subsequent classification.
Evaluation across three different motor imagery datasets demonstrates that our approach effectively addresses domain
adaptation challenges, outperforming both naïve cross-session and within-session methods. Our method eliminates the need
for data from new sessions, making it fully unsupervised concerning new session data and reducing the necessity for
recalibration with each session. Furthermore, the reduction of session-specific information in the reconstructed signals
indicates that our approach effectively denoises non-stationary signals, thereby enhancing the accuracy of BCI models. Future
applications could extend this model to a broader range of BCI tasks and explore the residual signals to investigate sources of
non-stationary brain components and other cognitive processes.
Go back