BCIwiki (bciwiki.com) — TongLiang Yang of Nanyang Technological University combined EEG and EMG signals for cross-user gesture recognition. ACF-Net achieved 87.6% accuracy on WAY-EEG-GAL and 89.2% on WEE-Ges, exceeding the strongest baselines by 9.4 and 9.1 percentage points. The paper appeared in the Journal of Neural Engineering on October 22, 2026.
Leave-one-subject-out validation tested each participant after training on the others. WAY-EEG-GAL included 12 participants; WEE-Ges included 20 participants over three days. Priors were constructed separately for each validation fold, and reported comparisons remained statistically significant after Holm-corrected Wilcoxon tests.
The network uses time-lagged coherence and phase-locking values to describe relationships between EEG and EMG channels. A Granger-causality-inspired attention module permits only EEG-to-EMG influence within a 30–120 ms window. This constrains information flow in the model; physiological causality requires independent evidence. Fused features are mapped to a symmetric positive definite manifold, a mathematical space used to compare gesture representations. Adaptation updates only the scale and offset parameters of batch-normalisation layers, using EEG preceding EMG as an unlabelled learning signal.
Ablation tests attributed gains of 6.3, 4.9, 3.4 and 1.8 percentage points to directed attention, prior injection, manifold embedding and test-time adaptation. Accuracy was 77.5% under 20 mm simulated electrode displacement. Adaptation reduced cross-day accuracy loss from 6.8 to 2.1 percentage points. On an NVIDIA Jetson Nano, a forward pass took 32.4 ms, with a 178 ms end-to-end response and operation at 30.9 Hz.
Participants were healthy young adults using gel-based or semi-dry wired hardware. Clinical populations, wireless dry electrodes and actual electrode movement during wear remain to be tested.