SADNet: sustained attention decoding in a driving task by self-attention convolutional neural network

Aim The lack of concentration is one of the primary causes of traffic accidents. Decoding attention states provides a way to monitor drivers' attention, thereby preventing tragedies. However, existing methods for decoding attention states from EEG signals face challenges such as insufficient feature extraction, inadequate representation of attention in a normal state, and weak interpretability.Methods To address these issues, we propose a sustained attention state decoding model called Sustain Attention State Decoding Neural Network (SADNet). By combining depthwise separable convolution and self-attention mechanisms, the model applies different attention to signals in the temporal and spatial domains, extracting effective local and global channel features for attention state recognition.Results In within-subject and cross-subject experiments on publicly available datasets, SADNet achieves state-of-the-art performance with an average F1-Score of 0.8894 and 0.6156 respectively, and an average AUC of 0.9545 and 0.7024, outperforming existing models in comparative experiments. Additionally, the results indicate that the SADNet model focuses on the central parietal, frontal, and anterior regions, identifying spindle waves as features for recognizing poor attention state, beta frequency for optimal attention state, and alpha frequency for normal attention state.Conclusion The proposed SADNet would be beneficial in general for sustained attentional decoding before stimulus-event or subject's responses.