Classification of EEG event-related potentials based on channel attention mechanism

Event-related potentials (ERPs) represent the electroencephalographic responses to specific stimuli and are crucial for analyzing and understanding the processing of conscious activities within the human brain. Their classification is of significant importance in psychology and cognitive science. To address the multichannel and high signal-to-noise ratio characteristics of EEG signals, we introduce a single-subject short-distance ERP superposition averaging method for preprocessing raw data and propose an ERP-Xception model that integrates an ECA module with depth-separable convolutions. The ECA module was modified to reduce potential information loss through hierarchical dimensionality reduction, effectively extracting channel weight information. The Xception architecture was optimized to minimize model parameters and inference time. Additionally, a feature panning module was incorporated in parallel, allowing for minor channel displacements to enhance model generalizability and robustness. Our model achieved the highest F1-scores of 74.7%, 84.5%, 81.2%, 50.6%, 93.5%, and 88.5% across six ERP datasets, including ERN, LRP, N2PC, N170, N400, and P3, thereby validating its effectiveness and transferability.