Identification of temporal variations in mental workload using locally-linear-embedding-based EEG feature reduction...

Area of Science:

• Neuroscience
• Human-Computer Interaction
• Machine Learning

Background:

• Accidents in safety-critical systems are often linked to human operator cognitive overload and inattention.
• Monitoring mental workload (MWL) variations over time is essential for preventing such incidents.
• Existing neuroimaging technologies offer potential for MWL assessment, but classification with limited data remains challenging.

Purpose of the Study:

• To develop and evaluate a novel EEG-based approach for classifying discrete mental workload (MWL) levels.
• To utilize representative MWL indicators and small training samples for accurate classification.
• To enhance safety in human-machine systems by enabling automatic detection of MWL variations.

Main Methods:

• Employing Locally Linear Embedding (LLE) to extract MWL indicators from different cortical regions.
• Utilizing Support Vector Clustering (SVC) to identify clusters representing binary MWL classes.
• Integrating Support Vector Data Description (SVDD) for detecting subtle indicator variations and refining cluster classification.
• Proposing a combined SVC-SVDD framework for automatic three-class MWL level identification (low, normal, high).

Main Results:

• The SVC approach successfully detected MWL variations, with clusters interpretable as binary MWL classes.
• The trained binary SVDD classifier demonstrated capability in identifying slight variations in MWL indicators.
• The integrated SVC-SVDD framework effectively classified three distinct MWL levels (low, normal, high) from experimental data.
• Comparative analysis showed the proposed framework achieves acceptable computational accuracy.

Conclusions:

• The proposed EEG-based SVC-SVDD framework offers an effective method for automatic MWL classification.
• This approach combines the strengths of both unsupervised and supervised learning strategies.
• The method demonstrates potential for improving operator safety in human-machine systems by monitoring cognitive states.
• The framework provides a robust solution for MWL assessment, even with limited training data.