EEG Based Mental Workload Assessment using Machine Learning

Abstract

Excessive mental workload effects mental and physical health along with the performance of individuals. There is a need to monitor the mental workload of operators performing critical tasks. Electrical signals produced by neural structures in the brain can be captured through EEG and information about mental state of an operator can be inferred. The power distribution in various frequency bands of these signals has been utilized to assess the mental workload. These noisy signals require significant filtering and preprocessing because of their low signal to noise ratio. A number of factors such as window size, filter cut-off, etc. influence the accuracy of EEG-based workload assessment. In this thesis, we analyze the performance of workload assessment pipeline with respect to these factors on an open-source workload dataset. Moreover, performance of workload assessment is analyzed using signals acquired from individual lobes of the brain instead of the entire brain and using different frequency bands instead of the entire frequency spectrum. Lastly, we also compare the performance of a number of classifiers for three level workload classification. For the preprocessing stage, a sliding window of 256 samples with an overlap of a quarter and an Artifact Subspace Reconstruction (ASR) threshold of 5 provide maximum assessment accuracy of 71.12%. Frontal and occipital lobes of the brain seem to contain the highest workload related information as they provide an average assessment accuracy of 65.83%. For frequency bands analysis, our findings validate that and bands are most relevant for workload assessment as they provide 72.23% assessment accuracy which is highest among all bands. Finally, Support Vector Machines (SVM) is able to classify mental workload with an average accuracy of 66.22% which is the highest among the classifiers compared.