Abstract
This paper presents a novel framework for mental state classification from electroencephalogram (EEG) data by modeling neural activity as a time series of graphs (TSG). Traditional methods rely on band frequency (BF) power, which often overlooks the dynamic relational structures between brain regions. Our proposed TSG approach induces connectivity graphs from EEG channels, computes temporal graph similarities, and utilizes spectral embeddings for feature extraction. Evaluation on stress (Mental Math) and cognitive load (MATB-II) datasets demonstrates that TSG features provide complementary information to BF features. Combining both modalities (TSG+BF) consistently achieves superior classification performance, reaching approximately 85% balanced accuracy in stress prediction tasks.
Introduction
EEG signals are characterized by high dimensionality and low signal-to-noise ratios, making the classification of "passive" mental states like cognitive workload or stress particularly challenging. Conventional feature extraction focuses on Band Frequency (BF) power—measuring the intensity of specific oscillations (e.g., Alpha, Beta) at individual sensors. While neuroscientifically grounded, BF features fail to capture the complex, time-varying interactions between disparate brain regions. This study introduces a graph-theoretic approach to automate the extraction of these relational features.
Methodology
The proposed pipeline consists of three primary stages: graph induction, temporal similarity analysis, and spectral embedding. As illustrated in Figure 1, EEG signals are first segmented into time windows. Within each window, channels are treated as nodes, and edges are defined by the Pearson correlation between signals, representing instantaneous functional connectivity. A sequence of these induced graphs is then analyzed to determine temporal dynamics, with similarity quantified using the Frobenius norm of the difference between correlation matrices.
To facilitate machine learning, these high-dimensional graph relationships are projected into a low-dimensional vector space using Adjacency Spectral Embedding (ASE) based on the Random Dot Product Graph (RDPG) model. An "out-of-sample" extension is implemented to allow real-time mapping of new EEG data into the learned embedding space without full re-computation, a critical requirement for practical Brain-Computer Interface (BCI) applications.
Datasets and Evaluation
The method was validated on two distinct datasets representing different classification challenges:
- Mental Math: 36 participants recorded during resting and stressful (recursive subtraction) states, representing a stress prediction task.
- MATB-II: 50 participants performing NASA’s Multi-Attribute Task Battery II, inducing four distinct levels of cognitive workload.
Results and Discussion
The primary finding of this research is the complementary nature of TSG and BF features. In the Mental Math task, graph-based features (TSG) outperformed frequency-based features (BF), while the reverse was observed in the MATB-II workload task. However, as shown in Figure 4, the combined feature set (TSG+BF) consistently yielded the highest balanced accuracy across both datasets and varying proportions of training data.
In the Mental Math dataset, the combined model reached ~85% accuracy with 80% training data, a significant improvement over BF alone (~72%). Furthermore, TSG features demonstrated robustness in transfer learning scenarios, particularly in "across-session" tests for MATB-II. Analysis of feature importance revealed that frontal and fronto-temporal connections were critical for stress classification, aligning with established neuroscientific models of stress processing. The TSG method also proved efficient, compressing complex EEG relationships into a tiny feature space (3 to 8 dimensions) while maintaining high predictive power.
Conclusion
Modeling EEG data as a time series of graphs provides a powerful, unsupervised method for capturing neural connectivity dynamics. By integrating these features with traditional frequency-based metrics, researchers can significantly improve the accuracy of mental state classifiers. Future work will explore more sophisticated graph similarity metrics and non-linear dependency measures, such as mutual information, to further refine the TSG framework.
Summary updated