“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,” explained study corresponding author Jing Jin from East China University of Science and Technology. The core innovations include (a) a fixed structure decoupler to separate category-related and independent features; (b) fine-grained patch coding and gated channel attention for spatiotemporal feature extraction; and (c) an Interclass Prototype Network (IPN) to enhance feature discriminability. “This hybrid approach enables the model to learn robust domain-invariant features without target domain data, significantly improving cross-subject generalization.”
The model leverages key technical advancements: The feature extractor uses multigranularity patch segmentation to capture multi-band EEG features and gated channel attention to focus on task-relevant brain regions. The domain-invariant feature module decouples features via four loss functions (classification, invariant feature learning, feature alignment, diversity promotion), while the IPN module optimizes feature distribution with cosine similarity metrics. “The synergistic design addresses EEG’s nonstationarity and high intraclass variance, ensuring both generalization and discriminability,” said co-author Junxian Li.
The study authors validated the model through extensive experiments on three public datasets (Giga, OpenBMI, BCIC-IV-2a): The DGIFE model achieved state-of-the-art accuracy across all datasets (77.36% on Giga, 84.08% on OpenBMI, 64.74% on BCIC-IV-2a) with low standard deviation, demonstrating stability. Ablation experiments confirmed the necessity of key modules—removing patch coding or channel attention reduced accuracy by 3-4 percentage points. The model also exhibited strong noise robustness, maintaining 69.20% accuracy at 0 dB SNR, outperforming baseline methods by 8-18 percentage points. Feature visualization verified alignment with neurophysiological principles (contralateral brain activation during 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,” said co-author Xiaochuan Pan. Future work will focus on adaptive hyperparameter optimization, dynamic patch size adjustment, and extension to more BCI paradigms (e.g., P300 speller). Overall, this domain generalization method provides a robust solution for cross-subject EEG decoding, advancing the practical application of BCI technology in medical rehabilitation, human-machine interaction, and other fields.
Authors of the paper include 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 under grant 62176090 and in part by Shanghai Municipal Science and Technology Major Project under grant 2021SHZDZX. This research is also supported by Project of Jiangsu Province Science and Technology Plan Special Fund in 2022 (Key research and development plan industry foresight, fundamental research fund for the central universities JKH01241605 and key core technologies) under grant BE2022064-1 and in part by the Lingang Laboratory under grant no. LGL8998.
The paper, “A Domain Generalization Method for EEG Based on Domain-Invariant Feature and Data Augmentation” was published in the journal Cyborg and Bionic Systems on Feb. 24, 2026, at DOI: 10.34133/cbsystems.0508.
The Lancet Oncology