BCIwiki (bciwiki.com) — A zero-calibration framework called RUNet decoded motor imagery EEG with an average cross-session accuracy of 87.19% on the BCI Competition IV 2a dataset, researchers reported in IEEE Transactions on Bio-Medical Engineering on January 12, 2026.
Variability in electroencephalography (EEG) signals across sessions and across subjects, driven by individual differences and environmental factors, is a major obstacle to neural decoding in brain-computer interface (BCI) applications. RUNet combines Riemannian manifold learning with unsupervised representation learning. A multi-scale spatiotemporal convolutional module captures local-global spatial features and multi-resolution temporal dynamics. A polysynergistic covariance optimization module applies multiple regularizations and adaptive fusion to counter EEG non-stationarity. A Riemannian Affine Log Mapping layer, built on Affine-Invariant Transformation and Log-Euclidean Mapping, runs end to end to reduce cross-domain covariance drift.
The model is trained with transfer learning. Pre-training applies an unsupervised contrastive loss to resting-state EEG to learn domain-invariant spatiotemporal features. Retraining then uses task-specific data to improve discriminability and feature disentanglement.
According to the abstract, RUNet reached average cross-session accuracies of 87.19%, 88.03% and 85.45% on the BCI Competition IV 2a and 2b datasets and on a self-collected laboratory dataset. Its cross-subject accuracies on the same datasets were 68.09%, 78.29% and 87.25%. On the PhysioNet dataset, it reached a cross-subject accuracy of 78.14%.