Precise Temporal P300 Detection in Brain Computer Interface EEG Signals Using a Long-Short Term Memory

Event-Related Potentials (ERP) detection is a latent problem in the clinical, neuroscience, and engineering fields. It is an open challenge that contributes to achieving more accurate and adaptable Brain-Computer Interfaces (BCI). The state-of-the-art typically uses simple classifiers based on Discriminant Analysis due to their little computational demand. Some more recent approaches have started using Deep Learning techniques, but these do not provide any temporal information and rarely focus on detecting the P300 at sample level in electroencephalography (EEG) signals, which would improve the Information Transfer Rate in BCIs. In other research areas, recurrent neural networks have shown high performance in those tasks that require online responses. We propose a new methodology, based on Long-Short Term Memory networks, in a sample level forecast to predict the P300 signal continuously. We get a slight improvement concerning the standard procedure, typically Bayesian Linear Discriminant Analysis, and we also show that the model predicts the occurrence of the P300 ERP at sample level in EEG signals. This brings us the possibility of evaluating the inherent variation between subjects. Our approach contributes to more agile and adaptable BCIs development, going further in the real-life usage of BCIs.