BCIwiki (bciwiki.com) — A two-stage multi-domain distillation network called TSMD-Net decoded motor intentions from the EEG of previously unseen subjects with mean accuracies of 70.21%, 77.82% and 79.53% on three public datasets, without using any target-subject data in training, researchers at the School of Automation, Hangzhou Dianzi University reported in Cognitive Neurodynamics on October 2, 2026 (PMID 42824656; DOI 10.1007/s11571-026-10564-7).
Motor imagery EEG decoding lets brain-computer interface (BCI) systems turn self-generated motor intentions into external control commands. Differences between individuals and the non-stationary nature of EEG cause distribution shifts that limit how well existing decoders generalize to new users. The authors argue that current domain generalization methods rely mostly on feature alignment or data augmentation and make little use of complementary discriminative knowledge among source subjects.
TSMD-Net trains multiple domain-specific heads and distills their knowledge into a unified leader head. Optimization proceeds in two stages: feature covariance alignment first establishes stable cross-domain representations, then multi-domain collaborative distillation with hyperbolic supervised contrastive learning improves cross-domain prediction consistency and class separation.
The team evaluated the method on two motor imagery benchmarks, BCI Competition IV 2a and 2b, and on the larger High Gamma Dataset for motor execution, in that order matching the three accuracy figures. The authors said the results show effective and generalizable cross-subject motor decoding across the three datasets.