Cross-subject EEG Decoding with spatiotemporal transformers for zero-calibration brain–computer interfaces

Brain–computer interfaces (BCIs) based on motor imagery (MI) provide a transformative direct communication pathway for patients with severe motor disabilities. However, the high inter-subject variability and non-stationarity of electroencephalography (EEG) signals necessitate time-consuming subject-specific calibration, which severely hinders the practical, “plug-and-play” deployment of BCI systems. To address this fundamental challenge, we propose BCI-Transformer, a novel spatiotemporal transformer framework specifically engineered for zero-calibration cross-subject EEG decoding. Our approach introduces localized spatiotemporal attention (LSTA) blocks that utilize neuroanatomical priors to capture stable, invariant neural signatures, coupled with a cross-subject manifold alignment (CSMA) loss to project diverse subjects into a shared, class-discriminative latent space. We conduct an exhaustive evaluation against 22 state-of-the-art baselines on three gold-standard benchmarks: BCI Competition IV 2a, 2b, and the large-scale OpenBMI dataset. Experimental results demonstrate that BCI-Transformer achieves a peak accuracy of 79.64% ± 9.18% on the 54-subject OpenBMI dataset, improving upon existing domain adaptation methods by up to 15.2% in zero-calibration accuracy. Additionally, multi-dimensional visualization of attention maps and feature grids reveals that the model effectively attends to physiologically relevant motor cortex regions (C3, C4, and Cz) across multiple frequency bands, providing high interpretability and biological validity for clinical applications. These findings demonstrate that combining anatomical priors with geometric alignment can substantially reduce calibration requirements, offering a practical pathway toward plug-and-play BCI systems that benefit patients with motor disabilities, clinicians in neurorehabilitation settings, and researchers developing assistive neural technologies. Future work will extend this framework to other BCI paradigms and real-time closed-loop deployment scenarios.