A novel AI-driven EEG generalized classification model for cross-subject and cross-scene analysis

Artificial intelligence algorithms combined with electroencephalography (EEG) can effectively identify and interpret patterns of brain activity. However, the considerable variability in EEG signals among individuals and the challenges in transferring data and features among different scenarios result in a lack of universality in EEG signal analysis methods. To address these challenges, we introduce a novel AI-driven EEG general classification model called the Deformation Residual Compact Shrinkage Attention Mechanism (D-RCSAM) network. This low- parameter model improves spatial sampling positions using deformable convolution blocks and reduces computational costs while improving generalization performance through depthwise separable residual blocks. We further optimized the soft thresholding function to enhance the model's nonlinearity and sparse representation, while also improving the loss function. We validated the proposed model on one public dataset and two private datasets, with results demonstrating that the D-RCSAM model effectively integrates both public and private EEG signal features. Visualization and interpretability results show that the D-RCSAM model can handle cross-subject and cross-scene classification tasks, outperforming state-of-the-art models in cognitive task classification. This research offers a new perspective on intelligent, comprehensive analysis across individuals and scenarios.