BCIwiki (bciwiki.com) — Researchers at Xi’an Jiaotong University, the University of Health and Rehabilitation Sciences and Zhengzhou University in China developed E-STACNet, reaching 88.00% mean accuracy on the BCI Competition IV 2a motor-imagery dataset. The paper appeared in Neural Networks on October 19, 2026.
The model processes spatial relationships between electrode channels and EEG changes over time in one network. Channel attention preserves relationships between recording locations, convolutional layers extract temporal and spatial features, and multi-head self-attention highlights informative time points before classification.
Tests covered three public datasets. Mean accuracies were 88.00% on BCI Competition IV 2a and 80.62% on Giga DB. On DEAP, valence and arousal classification reached 96.49% and 96.46%. Valence describes positive versus negative emotion, while arousal describes its activation level. The authors report better performance than the comparison methods on all three datasets.
Ablation tests supported the contribution of preserving channel relationships, while channel-importance visualizations offered clues about the information used by the model. Performance in practical use remains to be evaluated beyond these public datasets.