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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Fatigue significantly contributes to operational errors, accidents, and psychological issues.
• Accurate fatigue detection is hindered by challenges in identifying fatigue-related neural channels in electroencephalography (EEG) signals.
• Extracting effective features from high-dimensional EEG data for fatigue detection remains difficult.

Purpose of the Study:

• To propose the AXELLSTM fatigue detection framework for cross-subject EEG-based fatigue detection.
• To integrate interpretable feature selection with efficient temporal modeling for enhanced fatigue detection.
• To evaluate the framework's cross-subject generalization and reduce data leakage during feature and channel selection.

Main Methods:

• A nested leave-one-subject-out (LOSO) validation strategy was employed.
• ANOVA and XGBoost identified discriminative fatigue-related features, followed by L1-regularized sparse selection for channel importance.
• An EfficientLSTM model was developed, incorporating feature compression, bidirectional LSTM, channel attention, and temporal attention.

Main Results:

• Eight consistent features (e.g., mean, gamma power, variance) and 18 stable EEG channels were identified across all LOSO folds.
• The EfficientLSTM model achieved 0.7454 accuracy, 0.7436 balanced accuracy, and 0.7143 Macro-F1 under the LOSO protocol.
• EfficientLSTM demonstrated superior Macro-F1 compared to baseline models, balancing performance, complexity, and temporal modeling.

Conclusions:

• The AXELLSTM framework offers an interpretable and computationally efficient solution for EEG-based fatigue detection.
• The findings support the development of portable fatigue monitoring systems.
• The study highlights the distribution of fatigue-related EEG information across specific brain regions.