Theoretical and applied research on spatio-temporal graph attention networks for single-trial P300 detection

Objective.Accurate detection of single-trial P300 ERPs (event-related potentials) is crucial for developing high-performance non-invasive BCIs (brain-computer interfaces). However, this task remains challenging because of the low (signal-to-noise ratio) of EEG (electroencephalography) and the limited ability of existing models to concurrently capture the complex non-Euclidean spatiotemporal dynamics of brain signals.Approach.We propose a novel ST-GraphTRNet (spatiotemporal graph transformer network). This architecture synergistically integrates temporal convolutions for local feature extraction, graph convolutions to explicitly model the neurophysiological spatial relationships between EEG electrodes, and a temporal transformer with a self-attention mechanism to capture global, long-range temporal dependencies across the entire signal.Main results.Extensive evaluation of four public P300 datasets demonstrates that ST-GraphTRNet significantly outperforms (state-of-the-art) benchmarks under both within-subject and cross-subject paradigms. Crucially, interpretability analyzes via (T-distributed Stochastic neighbor embedding) and (Gradient-weighted Class Activation Mapping) revealed that the model's decisions aligned with established neurophysiological priors, focusing on parietal electrodes approximately 300 ms post-stimulus.Significance.This study provides a powerful and interpretable framework for single-trial ERPs decoding. By effectively integrating the strengths of (convolutional neural networks), (graph neural networks), and Transformers, a new benchmark for building high-accuracy, generalizable, and clinically viable BCIs is established, moving closer to the goal of plug-and-play systems that require minimal user-specific calibration.