Machine Learning Assessment of EEG Data in a fatigue related n-back task

Abstract

Mental fatigue significantly seriously impacts cognitive functions and decision-making, especially in safety-related environments like aviation, medicine, and transportation. This thesis presents a methodology for detecting mental fatigue from electroencephalography (EEG) signals recorded while doing an n-back working memory task. A hybrid deep learning model that combines a convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM) was developed to classify EEG signals as ”fatigued” and ”rested” states. In order to enhance the spatial resolution of the EEG signals, source localization was performed using the sLORETA algorithm, which projects scalp-recorded activity onto cortical surfaces. The model was trained and tested using 10-fold cross-validation, and achieved an average accuracy of 91.55%, demonstrating that it is robust and generalized across subjects. Explainability was also ensured through SHapley Additive exPlanations (SHAP), which provided insight regarding the most salient cortical sources that are responsible for the model predictions. The analysis highlighted contributions from frontal and parietal regions, that are consistent with neuroscientific findings on fatigue-related changes in brain activity. The dataset was collected from recordings of the participants in both rested and sleep-deprived conditions, enabling the model to learn discriminative patterns associated with mental fatigue. This work not only offers a high-performing and also explainable model for EEG-based fatigue assessment but also reduces the gap between deep learning and neuroscience by connecting machine learning predictions with physiologically meaningful brain processes.