Machine Learning Algorithms for Detecting Fatigue: An EEG Data Analysis

Abstract

Mental fatigue considerably affects cognitive performance, decision-making, and general productivity in diverse fields such as healthcare, transportation, and military operations. Extended cognitive strain may result in diminished vigilance and increasing error rates, presenting significant hazards to safety and productivity. Electroencephalography (EEG), a non-invasive technique for monitoring cerebral activity, offers an objective approach to identify and categorize mental fatigue. This thesis introduces a robust machine learning framework for classifying EEG data into rest and fatigue states, highlighting the application of specialized feature selection methods and machine learning classifiers. Electroencephalogram (EEG) data were obtained from 20 subjects performing 54 trials, categorized into rest and fatigue groups. The preprocessing procedures involved artifact elimination, including the removal of noise from muscular activity and eye blinks, as well as bandpass filtering into five frequency bands: delta, theta, alpha, beta, and gamma. Functional connectivity was assessed with the Phase Lag Index (PLI), a reliable metric of phase synchronization among EEG channels, producing high-dimensional datasets. To address the problem of dimensionality, eleven feature selection techniques, such as LASSO, ReliefF, Recursive Feature Elimination with Correlation Bias Reduction (RFE-CBR), and Fisher Score, were employed to discern the features that are most important while preserving interpretability. Five classifiers were trained using the chosen features: k-Nearest Neighbors (KNN), Support Vector Machine (SVM) with radial basis function (RBF) and linear

kernels, Linear Discriminant Analysis (LDA), and Random Forest (RF). Cross- validation methods were employed to guarantee the generalization of the chosen

features across different subjects. Performance was assessed utilizing criteria including accuracy, sensitivity, specificity, and F1-Score. The findings indicated that feature selection significantly enhanced classification performance. LASSO was identified as the most effective feature selection algorithm, achieving a combined accuracy of 97.5% with only 19 features and exhibiting excellent performance metrics (accuracy, sensitivity, specificity, and F1-Score) across all classifiers, demonstrating its efficacy in EEG-based fatigue detection. Lasso identified features that distinguish between rest and fatigue states using EEG channel connections. Connectivity was predominantly focused in the Frontal and Central lobes, indicating their functions in cognitive control and sensorimotor integration. Delta and Theta rhythms exhibited the highest differentiation, indicating their involvement in restorative processes and sustained attention under fatigue. These findings highlight LASSO's accuracy in identifying features relevant to fatigue identification. The present research illustrates that employing feature selection methods not only reduces the dimensionality of EEG data but also improves model interpretability by concentrating on the most prominent features. The suggested methodology establishes a basis for additional studies in EEG-based fatigue detection and presents possible applications in clinical and working environments, where the assessment of mental fatigue is essential for enhancing safety and performance.