Using Autoencoders to Denoise Cross-Session Non-Stationarity in EEG-Based Motor-Imagery Brain-Computer Interfaces

A major problem in brain-computer interfaces (BCIs) relates to the non-stationarity of brain signals. Consequently, the performance of a classification algorithm trained for an individual subject on a certain day deteriorates during the following days. The traditional approach is to recalibrate the algorithm every session, limiting the wide use of BCIs. Here, we use an autoencoder convolutional neural network to identify a low dimensional representation of the EEG signals from the first day (or days) and show that this allows for stable decoding performance on the following days without resorting to recalibration. Furthermore, we demonstrate that the residual signals, namely the difference between the original and reconstructed EEG, can be used to accurately discriminate among different recording sessions. In line with that, the reconstructed EEG cannot be used to discriminate among recording sessions. This implies that the reconstructed EEG reflects an invariant representation of the subject's intent, whereas the residual signals reflect a non-stationary component, which differs from one session to another. The findings are demonstrated through two different datasets.