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

Maintaining situational awareness and alertness is a significant challenge in high-stakes industries such as surveillance, transportation, security, etc. where even brief lapses in attention can have severe consequences. Although machine learning (ML) and deep learning (DL) have improved traditional methods of vigilance classification, the complexity of assessing human vigilance remains unsolved. Current approaches often rely on simplistic binary classification which often fails to capture the nuances of an operator's mental state. To advance our understanding, it is essential to develop more sophisticated models that can categorize vigilance into multiple levels, providing richer insights. Furthermore, the dynamic nature of human performance poses a significant challenge as it generates data of varying sizes. Existing solutions, such as data padding or truncation can introduce biases and compromise the accuracy of the results highlighting the need for more innovative and robust methods. This study proposes a graph embedding based approach that captures human vigilance subtleties through multi-level classification and effectively handles variable-sized EEG signals. Our experimental results demonstrate the superiority of Feather Graph Embedding (FG-Z_i) for EEG signals, achieving state-of-the-art performance for six-class vigilance classification. Specifically, our model attains an accuracy of 84.165% and F1-score of 83.734% on the training set, and 83.448% accuracy on the testing set, with a peak F1-score of 86.256%. Empirical findings indicate that the combination of FG-Z_i of EEG graphs with a 1D-CNN Multi-Headed Transformer framework holds great promise for accurate and real-time vigilance monitoring.Clinical relevance- The proposed graph embedding-based vigilance classification framework presents significant implications for both clinical and real-world applications, particularly in domains requiring continuous cognitive monitoring. By effectively modeling variable-length EEG signals and enabling multi-level vigilance assessment, the approach provides a more granular understanding of cognitive states. This advancement facilitates early detection of cognitive fatigue, enhances patient safety in clinical environments, and improves human-machine interactions in critical applications such as neuromonitoring, fatigue detection in medical professionals, and personalized cognitive training interventions.