An EEG-based fatigue detection framework integrating interpretable feature selection and efficient temporal modeling.

Introduction: Fatigue is a critical factor contributing to operational errors, traffic accidents, and psychological disorders. A major challenge in current fatigue detection lies in the difficulty of clearly identifying fatigue-related neural channels in the brain, which hinders the accurate extraction of effective features from high-dimensional electroencephalography (EEG) signals. Methods: This study proposes the AXELLSTM fatigue detection framework, an EEG-based cross-subject fatigue detection framework that integrates interpretable feature selection with efficient temporal modeling. Specifically, a nested leave-one-subject-out (LOSO) validation strategy was adopted to evaluate cross-subject generalization and reduce potential data leakage during feature and channel selection. Within each LOSO fold, ANOVA and XGBoost were first used to identify discriminative fatigue-related features, followed by an L1-regularized sparse selection model to estimate channel importance. Furthermore, an EfficientLSTM model was designed by integrating a feature compression module, bidirectional long short-term memory network, channel attention, and temporal attention mechanisms to capture sequence-level EEG temporal representations. Results: The results showed that eight features, including mean, RMS, zero-crossing rate, gamma power, beta power, standard deviation, variance, and peak-to-peak value, were consistently selected across all LOSO folds. In addition, 18 EEG channels, including F8, AF8, TP8, F2, Fp1, F4, FT10, F6, Fz, P7, FT9, O1, C5, FT8, CP4, F3, FC6, and P3, were stably retained, suggesting that fatigue-related EEG information is distributed across frontal, frontopolar, frontotemporal, parietal, and occipital regions. Under the LOSO protocol, EfficientLSTM achieved an accuracy of 0.7454 ± 0.2147, balanced accuracy of 0.7436 ± 0.2139, and Macro-F1 of 0.7143 ± 0.2452, with 0.1959 M parameters, a model size of 0.7474 MB, and 3.7069 M FLOPs. Compared with the baseline models, EfficientLSTM obtained the highest average Macro-F1 and showed a favorable balance between cross-subject recognition performance, model complexity, and temporal modeling capability. Discussion: These findings indicate that the proposed AXELLSTM framework provides an interpretable and computationally efficient solution for EEG-based fatigue detection and offers potential support for the development of portable fatigue monitoring systems.