Attention-based multi-semantic dynamical graph convolutional network for eeg-based fatigue detection

Liu, Haojie; Liu, Quan; Cai, Mincheng; Chen, Kun; Ma, Li; Meng, Wei; Zhou, Zude; Ai, Qingsong · 2023 · Crossref

DOI: 10.3389/fnins.2023.1275065

archive: archived pipeline: cataloged verified

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

Summary

This study addresses the critical need for reliable driving fatigue detection systems, noting that fatigue contributes to 20–30% of traffic accidents. While electroencephalogram (EEG) signals are considered the most direct physiological indicator of fatigue, traditional methods often overlook the functional connectivity between brain regions and fail to meet real-time processing requirements. Existing deep learning approaches, particularly Convolutional Neural Networks (CNNs), are limited by their reliance on Euclidean space, which hinders the modeling of non-Euclidean EEG data and inter-channel dependencies. To overcome these limitations, the authors propose the Attention-based Multi-Semantic Dynamical Graph Convolutional Network (AMD-GCN), a model designed to capture intrinsic topological relationships and focus on fatigue-relevant features. The AMD-GCN architecture comprises three primary modules: a Channel Attention Mechanism (AM-CAM), a Multi-Semantic Dynamical Graph Convolution (MD-GC), and a Spatial Attention Mechanism (AM-SAM). The model processes Differential Entropy (DE) features extracted from EEG signals across both 5 standard frequency bands and 25 finer 2-Hz resolution bands. The AM-CAM module assigns weights to input features to emphasize critical information. The core MD-GC module constructs dynamic adjacency matrices using three semantic patterns: spatial adjacency, Euclidean spatial distance, and self-attention mechanisms. This allows the network to adaptively learn the functional connectivity between EEG channels, capturing dependencies between both physically connected and distant nodes. Finally, the AM-SAM module removes redundant spatial node information to reduce interference. Experiments were conducted using the public SEED-VIG dataset, which contains EEG recordings from 23 participants engaged in simulated driving tasks. The dataset was annotated using PERCLOS (percentage of eye closure) metrics to classify states as awake, tired, or drowsy. The authors employed a subject-specific training protocol, preserving temporal order to prevent data leakage. The AMD-GCN model achieved a classification accuracy of 89.94%, surpassing existing algorithms. The results demonstrate that the proposed strategy effectively leverages multi-semantic graph structures and attention mechanisms to enhance the extraction of spatial features from EEG signals. The significance of this work lies in its effective integration of graph convolutional networks with attention mechanisms to address the non-Euclidean nature of EEG data. By dynamically constructing adjacency matrices based on multiple semantic patterns, the model captures complex functional brain connectivity that static graph methods miss. The high accuracy achieved on the SEED-VIG dataset validates the approach as a robust solution for real-time, EEG-based driving fatigue monitoring, offering a promising direction for improving road safety through advanced physiological signal processing.

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 qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

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

• drowsiness detection algorithms

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data