Functional Graph Image Representation applied to EEG-based Mental Workload Classification

Sarkis, Maria; Rizkallah, Mira; Moussaoui, Saïd · 2024 · Crossref

DOI: 10.1109/embc53108.2024.10781733

archive: archived pipeline: cataloged

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

Summary

This paper addresses the limitations of traditional functional connectivity metrics in EEG-based mental workload (MW) classification, specifically the issues of volume conduction and the neglect of electrode spatial locations. The authors propose a novel framework that learns a sparse functional graph from EEG data under structural constraints and converts it into a 2D image representation that explicitly encodes electrode positions, connection strengths, and node importance. This image is then processed by a convolutional neural network (CNN) to extract latent features for classification. The methodology involves three main steps: functional network construction, graph-to-image conversion, and CNN-based classification. First, a sparse graph Laplacian is learned from the EEG covariance matrix using a maximum a posteriori estimation of a Gaussian-Markov random field, subject to sparsity and structural constraints (either fully-connected or 3-nearest-neighbor). Second, the learned weight matrix is projected onto a 2D grid where pixel coordinates correspond to electrode locations; pixel values are determined by the maximum edge weight crossing that pixel, while electrode pixels reflect their degree (sum of incident edge weights). Third, these 61x61 images are fed into a CNN with three convolutional blocks (4, 8, and 16 filters) to predict MW classes. Experiments were conducted on a public dataset of 29 participants performing the MATB-II task at three difficulty levels (low, medium, high) across two sessions. The EEG signals were recorded from 64 electrodes (61 retained after preprocessing) at 250 Hz and decomposed into 2-second windows. The proposed framework outperformed state-of-the-art baselines, including Common Spatial Pattern with One-Versus-Rest (CSP-OVR) and Riemannian Minimum Distance to Mean (MDM). For instance, in Session 2 with an 80/20 train/test split, the proposed method achieved 90% accuracy with a fully-connected graph constraint, compared to 80% for Riemannian-MDM and 73% for thresholded Pearson correlation. The results demonstrate that learning sparse partial correlations yields substantial improvements over hand-crafted correlation metrics, and that imposing structural constraints reduces computational complexity without sacrificing performance. The significance of this work lies in its ability to integrate spatial and temporal aspects of EEG signals through a data-driven graph representation that preserves electrode geometry. By outperforming classical spatial filtering and Riemannian geometry approaches, the study validates the utility of graph image representations for EEG analysis. The findings suggest that exploiting the spatial layout of electrodes and sparse partial correlations enhances the discriminative power of deep learning models for cognitive state classification, offering a promising direction for future brain-computer interface applications.

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