Thesis

Deep learning classification model of mental workload levels using EEG signals

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2024
Thesis identifier
  • T16935
Person Identifier (Local)
  • 202057206
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Understanding and improving humance performance, especially in situations that require safety, productivity, and well-being, relies on categorising mental workload (MWL). Traditional methods for measuring MWL, such as in driving and piloting, have given us some understanding, but these methods must accurately distinguish between low and high workload levels. Excessive work can tyre participants, while insufficient work can make them bored and inefficient. Traditional MWL assessment tools, such as questionnaires, sometimes make it harder for people to manage their MWL, especially when they struggle to express or understand their thoughts and feelings. The recent work shift to neurophysiological signals, specifically electroencephalogram (EEG), provides a promising way to measure brain activity related to MWL non-invasively. Advanced techniques such as deep learning have made it easier to study EEG signals in more detail. Our goal was to develop a clear and consistent approach for using EEG signals to classify MWL effectively. Our approach focused on each process stage, from preparing the data to evaluating the model and addressing common mistakes and misunderstandings in current techniques. The first study addresses the challenges of using EEG data contaminated by artefacts for assessing MWL. EEG signal artefacts, such as eye movement or muscle activity, can skew MWL assessment. Recently, there has been significant progress in using deep learning models to interpret EEG signals, but the challenge remains. The preprocessing pipeline for EEG artefact removal is broad and inconsistently adopted; some pipelines are time-consuming and contain human intervention steps, so they are unsuitable for automation systems. Therefore, this study focused on automatic EEG artefact removal for deep learning analysis. Furthermore, we examined the impact of various preprocessing techniques on the effectiveness of deep learning models in classifying MWL levels. We used state-of-the-art models such as Stacked LSTM, BLSTM, and BLSTM-LSTM, and found that certain techniques—specifically, the ADJUST algorithm—significantly enhanced model performance. However, the sophisticated models could extract relevant information from raw data, indicating a reduced need for preprocessing. The second study shifted the focus to channel selection to refine the automation of MWL classification and reduce unnecessary computational expenses by using unnecessary electrodes, aligning more closely to real-world applications. We prioritised the best electrode setup focusing on brain activity related to MWL. We removed unnecessary data using Riemannian geometry, an effective method for EEG channel selection. We aimed to balance information sufficiency with computational efficiency and to reduce the number of electrodes. The study also evaluated covariance estimators for Riemannian geometry for their effectiveness in channel selection and impact on deep learning models for MWL classification, as the traditional Empirical Covariance (EC) has limitations for the EEG signal. Finally, the third study tackled a critical but frequently overlooked aspect of MWL level classification using machine learning or deep learning techniques: the temporal nature of EEG signals. We underscored that the traditional cross-validation technique violates the sequential nature of time series data, leading to data leakage, model overfitting, and inaccurate MWL assessment. Specifically, to predict the subject’s MWL level, we could not randomly split data and use future data to train the model and predict the previous MWL level. To address this problem, this study focused on the model training phase, specifically on the importance of time series cross-validation methods. We adopted the expanding window and rolling window strategies, finding that using the expanding window strategy outperformed those using the rolling window strategy. This research carefully developed a comprehensive and consistent method for classifying MWL using EEG signals. We aimed to correct misunderstandings and set a standard in brain-computer interface (BCI) systems. This will help guide future research and development efforts.
Advisor / supervisor
  • Moshfeghi, Yashar
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T16935 2024-05-21 Public Download