Adapting frozen foundation models for montage-agnostic high-resolution EEG event segmentation

Deploying brain-computer interfaces (BCIs) outside controlled laboratories requires detecting neural events in continuous electroencephalography (EEG) without relying on time-locked synchronization, while simultaneously generalizing across the diverse electrode montages encountered with different acquisition hardware. We investigate whether frozen EEG foundation models can be adapted to perform high temporal resolution event segmentation across unseen montages and datasets without subject-specific calibration. Approach. We introduce a lightweight, parameter-efficient preprocessing layer that interpolates learned channel embeddings based on electrode coordinates, enabling any frozen foundation model backbone to accept arbitrary montages. A shallow segmentation head is attached to produce a label every 4\,ms of continuous EEG, and overlapping predictions are consolidated via sliding-window majority voting. Main results. Evaluated on eight public corpora spanning P300, steady-state visually evoked potential (SSVEP) and motor imagery (MI) paradigms, our method consistently outperforms the original foundation models (BIOT, EEGPT) and classical baselines (EEGNet), achieving a mean macro F1 of 0.492 and Intersection over Union (IoU) of 0.361 in cross-subject evaluation, and F1\,=\,0.462, IoU\,=\,0.319 in calibration-free cross-dataset generalization. Significance. By decoupling the electrode montage from the pre-trained feature extractor through a plug-in adapter rather than massive retraining, our framework enables practical, resource-efficient BCI applications that operate without time-locked synchronization or montage-specific calibration, laying the groundwork for bridging the lab-to-field gap. The code and pre-processed datasets are available at: https://anonymous.4open.science/r/VewOdXnk669E17342jch-F1BD.