EEG attention decoding drops the fixed electrode graph: dynamic topology plus self-attention reaches 91.67% on three classes

EEG attention recognition no longer relies on a fixed electrode-connectivity prior; connection strengths between channels are updated during training according to feature similarity. Earlier graph convolution methods used a static adjacency matrix, treating electrode relations as a fixed structure that cannot represent brief, task-dependent coupling. In ablation comparisons, a static-topology graph convolutional network reached 84.8%, 6.9 percentage points below full DGCSA; dynamic graph convolution alone reached 87.9%, 3.8 percentage points below the full model.

Multi-band features are no longer concatenated directly; power spectral density and differential entropy receive separate weights per frequency band before fusion. Simple concatenation doubles feature dimensionality and introduces redundancy, and random forest and gradient boosting showed slightly lower accuracy after concatenation. Across SVM, random forest and gradient boosting, band-weighted fusion gave the highest accuracy, at 84.8%, 87.9% and 81.8% respectively.

The model adds the dynamic graph convolution and self-attention branches with equal weights, then passes the fused features through fully connected layers for low, medium and high attention classification. Convolutional networks model higher-order spatial dependencies poorly, long short-term memory networks favor temporal cues over channel structure, and multilayer perceptrons lack explicit structural modeling. Under identical features and data splits, DGCSA exceeded the multilayer perceptron at 85.61%, the convolutional network at 84.8% and the long short-term memory network at 87.88%.

The results rest on a self-collected dataset of 20 healthy university students aged 18 to 23, 64-channel acquisition and 873 three-second EEG segments, with attention labels defined by Schulte Grid completion time and accuracy. The applicable setting is three-class attention recognition in virtual and 2D learning tasks, plus comparison of attention scores between the two scenarios. The source proposes validation on larger, more diverse cohorts and the use of subject-independent protocols and external cohorts.

The source states that sample size and population diversity are limited, which may constrain generalization to other age groups and learner populations and may lead to overfitting. EEG acquisition in VR adds motion artifacts and higher sensory load; although preprocessing used visual inspection, semi-automatic independent component analysis labeling and manual confirmation, residual artifacts may still affect high-frequency features and learned connectivity patterns. The VR advantage over 2D may partly reflect a novelty effect, and randomized order with a one-day interval mitigates but does not eliminate it, so longitudinal repeated VR sessions are needed to determine whether the benefit persists after habituation. Low-attention samples misclassified as high attention are attributed to feature overlap between the two states and to labeling asymmetry between eyes-closed rest and task-engaged states.