BCIwiki (bciwiki.com) — According to the journal record from IEEE Transactions on Bio-Medical Engineering, a hybrid brain-computer interface (BCI) methods study by authors from Wuyi University, the University of Macau and the University of Essex appears in the November 2026 issue. The study proposes disentangled multimodal spatiotemporal learning (DMSL) to combine electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) signals, with evaluation on public datasets for mental arithmetic, motor imagery and emotion recognition.
The method addresses difficulties in capturing temporal and spatial relationships between the signals, as well as redundant information that can remain when each modality is learned as a whole. It first extracts spatiotemporal patterns from each modality and uses attention-based interaction to learn relationships between them. Graph convolution based on reconstructed channels, together with consistency and disparity constraints, then separates shared information from modality-specific information. These representations are adaptively fused for task prediction.
Table II reports mean accuracies of 75.43%, 68.24% and 45.40% for mental arithmetic, motor imagery and emotion recognition, exceeding the best task-specific baselines by 2.34, 0.59 and 1.47 percentage points, respectively. The comparison with the best baseline was statistically significant for mental arithmetic, but not for motor imagery or emotion recognition. These figures concern separate public-dataset tasks, not measurements of real-time performance. The authors interpret the results as supporting improved EEG-fNIRS decoding and strong generalization; that interpretation does not establish cross-session performance or clinical effectiveness.