ACF-Net: Design of an adaptive cross-modal fusion network for cross-subject wearable EEG-EMG gesture recognition.

Cross-subject gesture recognition in wearable brain-computer interfaces is constrained by four structural deficiencies: symmetric cross-modal attention discarding cortical precedence over muscular execution, priors confined to post hoc analysis, invariance anchored on unimodal features, and adaptation requiring labelled target-domain data. This study closes these gaps by embedding corticomuscular coupling directly into the network architecture rather than recovering it from data.
Approach: ACF-Net is a four-stage pipeline embedding neurophysiological priors directly into fusion. The corticomuscular coherence prior injection layer (CMC-Prior) encodes time-lagged coherence and phase-locking values for every electroencephalography (EEG)-electromyography (EMG) channel pair as a learnable bias tensor. Granger-causality-inspired directed cross-modal attention (GC-Attn) uses a causal mask allowing only EEG-to-EMG influence within 30-120 ms. The fused embedding is projected onto a symmetric positive definite manifold with log-Euclidean triplet loss. Physiology-consistent test-time adaptation (PC-TTA) updates only batch-normalisation affine parameters, using EEG-precedes-EMG as unlabelled self-supervision, eliminating labelled target data. Leave-one-subject-out validation on WAY-EEG-GAL (12 participants) and WEE-Ges (20 participants, three days) used fold-wise priors and Holm-corrected Wilcoxon tests.
Main results: ACF-Net reached 87.6% on WAY-EEG-GAL and 89.2% on WEE-Ges, exceeding the strongest baseline by 9.4 and 9.1 percentage points with all comparisons significant after correction. Ablation 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 remained 77.5% under 20 mm simulated electrode displacement, adaptation reduced the cross-day loss from 6.8 to 2.1 percentage points, and a forward pass on an NVIDIA Jetson Nano took 32.4 ms within a 178 ms end-to-end response at 30.9 Hz. 
Significance: Learned coupling weights recovered central-to-forearm-extensor projection, and delays matched corticospinal conduction with motor-unit recruitment, suggesting physiological structure, not dataset-specific statistics. Yet healthy young adults, gel-based/semi-dry wired hardware, and simulated displacement make this methodological; clinical and wireless dry-electrode transfer remains unestablished.