A Comparative Study of Classical Spectral and Spatial Feature Extraction Methods for Fatigue Detection via EEG Signals

Driving fatigue detection using electroencephalography (EEG) is an emerging application of smart systems with marked implication in transportation safety, healthcare and human-machine interaction. This study proposes a machine learning-based framework for binary mental fatigue classification, utilizing and comparing two established feature extraction methods-Power Spectral Density (PSD) and Common Spatial Pattern (CSP). A comparative evaluation was conducted across four traditional classifiers: k-Nearest Neighbors (kNN), Linear Support Vector Machine (SVM), Radial Basis Function (RBF) SVM, and Random Forest, using an open access EEG dataset. Four independent approaches were considered for this study: PSD-based features with all features considered, feature selection of the most informative PSD features, CSP features with all components considered, and selected CSP components. Each approach considered utilizing all four classifiers. Experimental results revealed that in general, CSP-based spatial features outperformed PSD in three out of four classifiers, specifically in non-linear models. Notably, RBF SVM with selected CSP components yielded an accruacy of 91.13%, comparable to kNN with all CSP components (91.63%) while outperforming both PSD-based approaches. These findings highlight the effectiveness of CSP-based spatial filtering combined with ML for EEGsignal classification representing a promising step towards the development of real-time, intelligent fatigue monitoring systems.