EXPLOITING SPATIOTEMPORAL EEG STRUCTURE USING DEEP LEARNING FOR IMPROVED P300 BRAIN–COMPUTER INTERFACES

Brain Computer Interfaces (BCIs) enable communication for individuals with severe motor impairments, with the P300 speller being a widely used paradigm that relies on detecting the P300 event-related potential in EEG signals. Traditionally, state-of-art approaches use 2D EEG representations, stacking signals from multiple channels into a channels-time matrix. While this captures temporal and channel-wise features, it ignores spatial relationships among electrodes. This work aims to improve P300 detection and character recognition by explicitly modeling both spatial and temporal information in EEG data. In this work, we propose AttenPCA-Net-GRU; a deep learning model that integrates convolutional layers, squeeze-and-excitation blocks, and spatial attention to capture discriminative spatial features. EEG signals are first preprocessed and transformed into a spatially informed 3D representation by mapping electrode signals onto a scalp-based grid. Principal Component Analysis (PCA) is then applied per channel, and the resulting components are organized as sequential spatial frames. To further incorporate temporal dynamics across PCA components, we include a Gated Recurrent Unit (GRU) branch that models dependencies across the frame sequence. AttenPCA-Net-GRU is evaluated on the BCI Competition III Dataset II P300 speller benchmark dataset. Preliminary results show that AttenPCA-Net-GRU achieves higher character recognition accuracy across different numbers of trials, with improvements of up to 3% compared to AttenPCA-Net (i.e., without the GRU branch), and performance increases as the number of trials increases. These findings demonstrate that jointly exploiting spatial electrode relationships and temporal structure enhances P300 speller performance and supports the development of more accurate and reliable EEG-based BCIs.