Graph Force Learning from EEG Signals for Mental Workload Classification

Sarkis, Maria; Rizkallah, Mira; Moussaoui, Saïd · 2025 · Crossref

DOI: 10.1109/ssp64130.2025.11073323

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

**Research Question and Motivation** This paper addresses the challenge of effectively representing electroencephalography (EEG) signals as graphs for mental workload classification. While graph-based frameworks have gained traction in neuroscience, existing methods often under-explore the spatial positioning of electrodes when integrating structural and statistical features. The authors propose a novel "Graph Force Learning" framework that leverages both the physical spatial arrangement of electrodes and a data-driven learned graph structure to define interaction forces between nodes, aiming to create a more discriminative feature representation than conventional approaches. **Methods and Experimental Design** The proposed method involves three main steps: learning a graph Laplacian from EEG data under sparsity and connectivity constraints; computing attractive and repulsive forces between nodes based on physics-inspired models (where attractive forces depend on connection strength and distance, and repulsive forces depend on node degrees and inverse squared distance); and extracting force-related node features (either Cartesian components or magnitude/orientation) to form a single-vector graph embedding. These features are then classified using Support Vector Machines (SVM). The approach was evaluated on the "COG-BCI" dataset, specifically using MATB-II recordings from 29 participants across two sessions, involving three difficulty levels (low, medium, high). EEG signals from 61 electrodes were segmented into 2-second overlapping windows. The learned Laplacian was derived using a regularization parameter $\alpha = 0.01$. Performance was assessed via 5-fold cross-validation, with hyperparameters optimized through grid search. **Results** The study compared different graph construction strategies and feature extraction methods. Learning the graph Laplacian from EEG signals yielded superior performance compared to hand-crafted functional connectivity measures like Pearson Correlation or Phase Locking Value. Specifically, the learned Laplacian with Cartesian force components $[F_u, F_v]$ achieved an average accuracy of $0.90 \pm 0.08$, whereas Pearson Correlation achieved $0.77 \pm 0.08$ and Phase Locking Value achieved $0.66 \pm 0.07$. Both Cartesian components and magnitude/orientation representations performed similarly, with Cartesian components showing a slight advantage (~2%). When compared against conventional baselines, the proposed framework outperformed Riemannian Minimum Distance to Mean (MDM) on covariance matrices ($0.62 \pm 0.08$) and Common Spatial Pattern (CSP) with One-Versus-Rest (OVR) SVM ($0.82 \pm 0.12$). **Significance** The results demonstrate that incorporating physics-inspired force interactions, which explicitly account for electrode spatial positions and learned connectivity, provides a more informative feature space for mental workload classification than traditional spatial domain methods. This approach effectively bridges the gap between structural graph properties and signal statistics, offering a robust alternative to deep graph neural networks for EEG analysis. The findings suggest that force-based embeddings can enhance classification accuracy in cognitive state monitoring, with future work directed toward higher-order graph representations and interpretability.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices