Neural Noise- Ultra-high-density EEG (uHD EEG) is used for individual finger decoding; accurate classification is difficult because of subtle spatial overlap in cortical activity.1
- A Sandwich enhanced Convolutional Block Attention Module (SCBAM) integrates dual attention between convolutional layers to improve individual finger classification from uHD EEG.1 1
Weekly enrichment (2026-07-20)
- The SCBAM (“Sandwich enhanced CBAM”) architecture places convolutional layers before and after a Convolutional Block Attention Module, whose two sequential attention mechanisms handle filter-domain then spatial/rhythm-domain information; inputs are FBCSP features rather than raw EEG.23
- In binary classification across all ten finger pairs in five subjects, SCBAM reached an average accuracy of 78.63% (±1.56), with the highest pairwise accuracy of 85% on thumb-vs-ring.2
- For five-class individual-finger classification, SCBAM averaged 61.12% (±0.95) accuracy, peaking at 62.36% for subject S2, built from ten one-vs-one binary classifiers aggregated by majority voting.2
- The authors state this is the first attempt at five-class individual-finger decoding using ultra-high-density EEG (>256 scalp electrodes), addressing a gap where prior work stopped at binary pairs.23
- SCBAM outperformed reported baselines on comparable uHD EEG data, including SVM at 64.8% (Lee et al., 2022) and MLP at 65.68% (Nemes et al., 2024), as well as EEGNet and five other benchmark networks using FBCSP features.2
- Training used 150 epochs, batch size 32, learning rate 0.001 and dropout 0.3; each 5 s trial (3000 samples) was chopped into 600-sample windows with 50% overlap, yielding 2,250 trials.2
- Data hygiene: Common Average Reference preprocessing, runs 1–8 used for training and runs 9–10 held out for test, with FBCSP reducing data volume by roughly 28% and trial-wise labels assigned by majority vote over nine windows.2
- Ablation studies confirmed that both sequential attention modules contribute, and that adding more convolutional layers around the CBAM did not improve accuracy.2
- The results support ultra-high-density EEG as a feasible non-invasive route to dexterous finger decoding for brain–computer interfaces, avoiding the craniotomy required for ECoG.23