Enhanced P300-based brain-computer interface decoding via region-weighted multiscale temporal-spatial representation learning method

P300-based brain-computer interface (BCI) systems provide a vital communication channel for individuals with severe physical disabilities. While deep learning has achieved significant progress in signal processing, P300 recognition remains hindered by low signal-to-noise ratios and non-stationary dynamics. Existing deep learning approaches exhibit two primary limitations: they rely on fixed-scale temporal filters that fail to capture multi-resolution event-related potential (ERP) dynamics, and they treat electrode channels homogeneously, overlooking the functional modularity of cortical regions and the interaction between local and global network patterns. To address the above issue, we propose the Region-Weighted Multiscale Temporal-Spatial Neural Network (RWMTS-Net). The model integrates three key components: a region partition module that groups physiologically aligned electrodes, a multiscale temporal-spatial module that captures multi-resolution ERP dynamics across functional brain regions, and an adaptive channel attention mechanism that dynamically couples local dynamics with global network features. Evaluated on a public dataset and a self-collected dataset under participant-dependent conditions, RWMTS-Net significantly surpassed state-of-the-art baselines (including EEGNet, ShallowNet, EEGInception, PPNN, and DeepNet). It achieved superior average character recognition accuracies of 98.93% and 91.79% on the respective datasets, along with enhanced information transfer rates (ITRs), higher AUC, and lower performance variance. These results demonstrate that RWMTS-Net effectively resolves key spatio-temporal representation bottlenecks, offering a highly accurate and robust solution for BCI decoding.