Non-invasive BCI

A multidisciplinary expert panel developed a consensus on functional reconstruction for post-stroke spastic hemiplegia. It recommends distinguishing impaired voluntary motor control, abnormal muscle overactivity, fixed structural restrictions, and muscle force imbalance before selecting treatment according to functional goals and the main limiting factor. Task-oriented motor training remains fundamental when voluntary control is impaired, while non-invasive BCI-based closed-loop training may be an adjunct for selected patients. The consensus also addresses diagnostic motor nerve block, peripheral nerve surgery, and tendon and soft-tissue reconstruction. Because evidence quality varies, no universal treatment sequence or combination strategy is currently supported.

Researchers from Kocaeli University and collaborating institutions proposed an adaptive multi-domain feature-fusion framework to distinguish left-hand motor imagery, right-hand motor imagery, mental arithmetic and rest from functional near-infrared spectroscopy signals. Using subject-independent leave-one-subject-out cross-validation on an open-access dataset of 29 participants, the framework achieved 95.6% mean accuracy, compared with 90.7% for temporal features alone, 95.0% for cepstral features alone and 95.2% for direct concatenation. Accuracy gains over direct concatenation ranged from 0.2 to 0.7 percentage points for every participant. Most of the gain over temporal features came from the cepstral domain, with adaptive weighting adding a small, consistent improvement.

A collaboration involving Nanchang University proposed a deep learning network to classify motor imagery tasks from EEG signals. Across three public BCI datasets, it achieved average accuracies of 80.21% for Dataset 1's 4-class tasks, 90.19% for Dataset 2's 2-class tasks, and 97.32% for Dataset 3's 4-class tasks, exceeding the strongest comparison network on each dataset. The approach combines multiscale processing with temporal and channel attention, while augmenting training data through segmentation and reconstruction. These results reflect classification under the tested dataset and training conditions, not demonstrated real-time BCI or clinical performance.

A pilot study involving authors from Air University and other institutions recorded functional near-infrared spectroscopy signals from 20 healthy participants and 7 patients with ankle motor impairment. It sought to distinguish rest, dorsiflexion and plantar flexion and generate exo-ankle control commands. In cross-participant offline validation among healthy participants, a support vector machine achieved a mean accuracy of 72.50% ± 13.8%. Simulated-online testing with participant-specific validation reached its highest mean accuracy of 62.2% ± 14.1% at a 1.5-s step size. The patient group achieved 57.23% ± 5% under leave-one-trial-out validation. These are proof-of-concept results, not demonstrations of prospective real-time exo-ankle control or rehabilitation efficacy.

A study in healthy adults compared flickering disks filled with Gabor textures against plain flickering disks. An initial frequency sweep involved 24 participants; analysis of a five-target task two months later included 18 of them. Textured flicker produced higher EEG signal-to-noise ratios at 3-9 Hz and was rated more comfortable and less fatiguing at every tested frequency. In offline classification of the five-target task, textured flicker performed well at 5-7 Hz, while plain flicker was more reliable at 14-16 Hz. Accuracy was comparable within each stimulus's preferred band. Better comfort therefore did not mean better decoding at every frequency.