“Domain bias caused by individual differences and device variations severely limits BCI’s practical application, while existing methods struggle with feature decoupling and noise sensitivity,” defined examine corresponding creator Jing Jin from East China University of Science and Technology. The core improvements embody (a) a hard and fast construction decoupler to separate category-related and impartial options; (b) fine-grained patch coding and gated channel consideration for spatiotemporal characteristic extraction; and (c) an Interclass Prototype Network (IPN) to reinforce characteristic discriminability. “This hybrid approach enables the model to learn robust domain-invariant features without target domain data, significantly improving cross-subject generalization.”
The mannequin leverages key technical developments: The characteristic extractor makes use of multigranularity patch segmentation to seize multi-band EEG options and gated channel consideration to give attention to task-relevant mind areas. The domain-invariant characteristic module decouples options via 4 loss capabilities (classification, invariant characteristic studying, characteristic alignment, variety promotion), whereas the IPN module optimizes characteristic distribution with cosine similarity metrics. “The synergistic design addresses EEG’s nonstationarity and high intraclass variance, ensuring both generalization and discriminability,” stated co-author Junxian Li.
The examine authors validated the mannequin by means of in depth experiments on three public datasets (Giga, OpenBMI, BCIC-IV-2a): The DGIFE mannequin achieved state-of-the-art accuracy throughout all datasets (77.36% on Giga, 84.08% on OpenBMI, 64.74% on BCIC-IV-2a) with low normal deviation, demonstrating stability. Ablation experiments confirmed the need of key modules—eradicating patch coding or channel consideration decreased accuracy by 3-4 share factors. The mannequin additionally exhibited robust noise robustness, sustaining 69.20% accuracy at 0 dB SNR, outperforming baseline strategies by 8-18 share factors. Feature visualization verified alignment with neurophysiological ideas (contralateral mind activation throughout motor imagery).
“While the DGIFE model shows strong performance, it faces limitations: sensitivity to hyperparameters such as temperature coefficients, and reliance on predefined patch lengths,” stated co-author Xiaochuan Pan. Future work will give attention to adaptive hyperparameter optimization, dynamic patch dimension adjustment, and extension to extra BCI paradigms (e.g., P300 speller). Overall, this area generalization technique supplies a sturdy answer for cross-subject EEG decoding, advancing the sensible utility of BCI know-how in medical rehabilitation, human-machine interplay, and different fields.
Authors of the paper embody Jing Jin, Junxian Li, Xiaochuan Pan, Ren Xu, Andrzej Cichocki, Wenli Du, and Feng Qian.
This work was supported by Brain Science and Brain-like Intelligence Technology-National Science and Technology Major Project 2022ZD0208900 and National Natural Science Foundation of China below grant 62176090 and partially by Shanghai Municipal Science and Technology Major Project below grant 2021SHZDZX. This analysis can also be supported by Project of Jiangsu Province Science and Technology Plan Special Fund in 2022 (Key analysis and improvement plan trade foresight, elementary analysis fund for the central universities JKH01241605 and key core applied sciences) below grant BE2022064-1 and partially by the Lingang Laboratory below grant no. LGL8998.
The paper, “A Domain Generalization Method for EEG Based on Domain-Invariant Feature and Data Augmentation” was revealed within the journal Cyborg and Bionic Systems on Feb. 24, 2026, at DOI: 10.34133/cbsystems.0508.