SMFF-Net: Spatiotemporal-frequency Multi-domain Feature Fusion Network for EEG-based brain state detection

The timely and reliable detection of driver fatigue is crucial for reducing driving risks and improving traffic safety. EEG-based deep learning approaches for brain state detection suffer from insufficient cross-domain feature interaction and limited neurophysiological interpretability. In this work, we propose a Spatiotemporal-frequency Multi-domain Feature Fusion Network (SMFF-Net) that characterizes fatigue-related EEG representations by the coordination of temporal, spatial and frequency features. SMFF-Net utilizes three complementary branches that leverage Convolutional Neural Networks (CNNs) and Graph Neural Networks (GCNs) to extract multi-domain features of EEG signals. An Enhanced Cognitive Importance Modulation (ECIM) module is designed to optimize domain-specific discriminative features, and an Entropy-Guided Gated Fusion (EGGF) module is proposed to adaptively integrate the enhanced multi-domain features by jointly considering branch-wise uncertainty and cross-domain complementarity. Experiments on SEED-VIG and SADT datasets show that SMFF-Net achieves superior performance over state-of-the-art methods, with accuracies of 98.54% and 99.74% and F1-scores of 98.39% and 99.73%, respectively. Model-derived brain activity analysis further indicates that the discriminative features of SMFF-Net are associated with fatigue-related EEG representations, characterized by the destabilization of temporal dynamics, the reconfiguration of inter-regional brain connectivity, and the slowness of spectral rhythm. The results indicate that coordinated multi-domain representation modeling enables SMFF-Net to improve the effectiveness of EEG-based driver fatigue detection while providing a neurophysiologically interpretable characterization of brain state alterations.