Researchers built an adapter that lets EEG foundation models read brain signals from electrode layouts they were never trained on.
The tool, called CortexBridge, combines raw EEG features with electrode and brain-atlas coordinates to map any montage into a shared cortical latent space. The team tested it on three frozen foundation models (EEGPT, LaBraM, and CBraMod) across five brain-computer interface datasets from the MOABB benchmark suite. CortexBridge improved performance in 13 of 15 evaluations, with average balanced-accuracy gains of 0.80% for EEGPT, 0.70% for LaBraM, and 3.26% for CBraMod. The single biggest jump was a 13.02% gain on a 12-class steady-state visual evoked potential classification task.
EEG foundation models are usually locked to the electrode layout they were trained on, which is a problem since every lab, clinic, and headset uses a different cap. An adapter that translates between montages without retraining the whole model could make it far easier to reuse these models across different hardware and settings, especially for brain-computer interfaces built outside the lab that generated the training data.
The gains are real but modest for two of the three models, and a single 13% jump on one niche visual task is doing a lot of the heavy lifting for that headline average.