MSDA-Net: Multiscale Spatiotemporal Dual-Attention Network for EEG-Based Driver Fatigue Detection

Driver fatigue poses a severe risk to road safety, contributing to approximately 20% of fatal accidents worldwide. While EEG signals are the gold standard for detecting fatigue, existing methods struggle to capture the complex spatiotemporal patterns in EEG data. We propose MSDA-Net, a multiscale spatiotemporal dual-attention network that integrates multiscale CNNs, GRUs and dual-attention mechanisms to dynamically prioritise spatial channels and temporal segments, which are critical for fatigue detection. The model processes EEG data through three blocks: a multidimensional signal encoding block, which transforms raw signals into 4D differential entropy features; a multiscale spatial attention block, which extracts local and global spatial patterns; and a temporal modelling block, featuring a GRU and temporal attention. Finally, a fully connected layer and sigmoid activation are used to classify fatigue states. Evaluated on the SEED-VIG dataset, MSDA-Net achieves state-of-the-art performance, significantly outperforming existing methods. This study can provide new insights into brain fatigue research and play a significant role in advancing the field's development.