ATCRN: Attention-guided Temporal Convolutional Remix Network for P300 speller.

BACKGROUND: The P300 speller is a prominent brain-computer interface (BCI) that facilitates communication by detecting P300 event-related potentials. However, its performance is substantially constrained by the low signal-to-noise ratio of EEG signals and the inherent temporal variability of the P300 response. NEW METHOD: We propose the Attention-guided Temporal Convolutional Remix Network (ATCRN), an end-to-end model that synergistically integrates a novel Temporal Convolutional Remix Network (TCRN) with a dual-attention framework. The TCRN employs multi-level skip connections to enable dynamic, cross-hierarchical fusion of local and global temporal features, addressing the variable latency of P300. Externally, the Convolutional Block Attention Module (CBAM) suppresses noise in spatial and channel dimensions. Internally, Efficient Channel Attention (ECA) modules within TCRN block perform dynamic channel recalibration. RESULTS: On BCI Competition III Dataset II, ATCRN achieved character recognition rates of 99% and 98% for two subjects at the 15th repetition, and yielded superior information transfer rates. Evaluation across eight ALS patients showed robust P300 detection (average AUC-ROC 0.882). COMPARISON WITH EXISTING METHODS: ATCRN outperforms both established CNN/TCN benchmarks and recent Transformer-based models across two public datasets, achieving state-of-the-art results in P300 detection and character spelling. CONCLUSION: The proposed ATCRN provides a novel, robust, and effective decoding framework for the P300 speller. The integration of TCRN for temporal feature fusion and the dual-attention mechanism for feature refinement offers a practical solution for advancing BCI applications.