Vigilance Classification for Variable-length EEG Signals using Graph Projections & Transformers.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Maintaining situational awareness is critical in high-stakes industries, but current vigilance classification methods are often too simplistic.
• Existing machine learning and deep learning models struggle with the dynamic nature of human performance data, leading to biases and inaccuracies.
• There is a need for sophisticated models capable of multi-level vigilance classification and robust handling of variable-sized electroencephalogram (EEG) signals.

Purpose of the Study:

• To develop an advanced graph embedding-based approach for multi-level vigilance classification.
• To effectively manage variable-sized EEG signals without introducing data biases.
• To enhance the accuracy and granularity of cognitive state assessment for improved vigilance monitoring.

Main Methods:

• Proposed a novel Feather Graph Embedding (FG-Z_i) approach for EEG signals.
• Utilized a 1D-CNN Multi-Headed Transformer framework for vigilance classification.
• Implemented multi-level classification to capture nuanced mental states.

Main Results:

• Achieved state-of-the-art performance for six-class vigilance classification using FG-Z_i on EEG data.
• Attained 84.165% accuracy and 83.734% F1-score on the training set.
• Reached 83.448% accuracy and a peak F1-score of 86.256% on the testing set.

Conclusions:

• The proposed graph embedding-based framework offers a promising solution for accurate, real-time vigilance monitoring.
• This method provides a more granular understanding of cognitive states, aiding in early detection of fatigue.
• Significant implications for clinical and real-world applications, including neuromonitoring and cognitive training.