BCIwiki (bciwiki.com) — A lightweight adapter called CortexBridge maps arbitrary EEG electrode layouts into a shared cortical latent space and improved 3 frozen foundation models in 13 of 15 evaluations across 5 brain-computer interface (BCI) datasets, according to a preprint posted to arXiv on October 1, 2026. The work by Jiazhen Hong, Xiaotian Zhou, Zihao Ding, Kailong Wang and Yu Wu has not been peer reviewed.
The authors note that EEG foundation models are often pretrained with a fixed channel vocabulary or a limited set of montages, which makes transfer difficult when electrode layouts change. CortexBridge combines EEG features with electrode and atlas coordinates to route signals into a common cortical space before they reach the pretrained model.
The 5 datasets come from the Mother of All BCI Benchmarks (MOABB). Gains in balanced accuracy averaged 0.80% for EEGPT, 0.70% for LaBraM and 3.26% for CBraMod, with a maximum gain of 13.02% on 12-class steady-state visual evoked potential (SSVEP) classification.
Visualizations of the learned atlas representations showed task-dependent spatial patterns, with SSVEP more concentrated in the Yeo Visual network than auditory P300. The authors say the results establish cortical alignment as a learnable, anatomically grounded way to connect heterogeneous EEG montages to pretrained foundation models.