Signal, Image and Video Processing Обработка сигналов
Enhancing inter-subject generalization in driver fatigue detection via energy-informed ICLabel EEG preprocessing
Аннотация
Electroencephalography (EEG) is widely used in cognitive state decoding and driver fatigue detection, yet its robustness remains limited by inter-subject variability and the challenge of automatically removing non-brain components. A common strategy combines Independent Component Analysis (ICA) with ICLabel to identify and reject artifacts. However, ranking non-brain components based solely on ICLabel’s probabilistic scores can result in misclassification and degrade the quality of cleaned EEG signals. Unlike conventional pipelines that rely solely on ICLabel’s probabilistic outputs, our method introduces an energy-based criterion to identify and rank truly harmful components. By assessing each component’s energetic impact on noisy channels, it selectively removes those most detrimental to neural features while preserving task-relevant activity. This balance enhances both EEG quality and cross-subject fatigue detection performance under strict validation protocols (LOSO-CV and LODO). We evaluated the proposed energy-informed preprocessing on two public driver-fatigue EEG datasets under strict LOSO and LODO validation with lightweight classifiers: Random Forest (RF), Support Vector Machine with RBF kernel (SVM), and Extreme Gradient Boosting (XGB). For fair comparison, all methods used identical features and classifiers, differing only in preprocessing. The proposed approach achieved up to 88.1% balanced accuracy and 94.3% fatigue sensitivity on DS1, with same-classifier improvements ranging from +1.9 to +12.3 points on DS2 LOSO and consistent superiority across all six classifier-dataset combinations. Using only Welch PSD features with simple classifiers, the performance was robust and reproducible, indicating strong cross-subject and cross-dataset generalization while remaining computationally simple for real-world deployment.