High-accuracy EEG-based driver fatigue detection using feature-coupled tensor decomposition for multimodal brain...

Driver fatigue is a major threat to road safety, and electroencephalography (EEG)-based monitoring offers an effective approach for assessing a driver's cognitive state. However, existing brain network feature-extraction methods often rely on single-modality connectivity metrics, which may not fully capture the complex multidimensional reorganization of neural dynamics during the transition to fatigue. To address this methodological gap, this study proposes a Feature-Coupled Tensor Decomposition (FCTD) framework for multimodal brain network analysis. The framework constructs independent high-order tensors for three complementary connectivity metrics: signal covariance (representing amplitude coupling), phase-locking value (representing phase synchronization), and Liang-Kleeman information flow (representing directed information transfer). Using Tucker decomposition, FCTD learns spatial bases at the scalp-electrode level and employs a cross-mapping strategy to integrate latent inter-modality dependencies. The proposed framework was evaluated on two publicly available driver fatigue EEG datasets. Under a chronological block-wise 10-fold cross-validation protocol designed to reduce the risk of temporal leakage, the method achieved average accuracies of 99.42% and 93.62% on Dataset 1 and Dataset 2, respectively. Leave-one-subject-out (LOSO) cross-validation on Dataset 2 yielded an accuracy of 75.18%, indicating that cross-subject generalization remains challenging under substantial inter-subject variability. By integrating complementary connectivity views, this study provides a computationally efficient framework with feature projections for multimodal EEG feature extraction in fatigue monitoring systems.