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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Deploying brain-computer interfaces (BCIs) outside labs requires robust neural event detection in electroencephalography (EEG) despite varying electrode montages and without time-locked synchronization.
• Existing EEG foundation models often struggle with generalization across different hardware and datasets, necessitating subject-specific calibration.

Purpose of the Study:

• To investigate the adaptability of frozen EEG foundation models for high-temporal-resolution event segmentation in continuous EEG.
• To enable generalization across diverse electrode montages and datasets without requiring subject-specific calibration.
• To develop a practical and resource-efficient BCI framework for lab-to-field deployment.

Main Methods:

• Introduced a lightweight, parameter-efficient preprocessing layer for interpolating channel embeddings based on electrode coordinates.
• Enabled frozen foundation models to accept arbitrary EEG montages via a plug-in adapter.
• Attached a shallow segmentation head for 4ms temporal resolution labeling and used sliding-window majority voting for prediction consolidation.

Main Results:

• Achieved a mean macro F1 of 0.492 and IoU of 0.361 in cross-subject evaluation across eight public corpora (P300, SSVEP, MI paradigms).
• Demonstrated calibration-free cross-dataset generalization with F1=0.462 and IoU=0.319.
• Outperformed original foundation models (BIOT, EEGPT) and classical baselines (EEGNet).

Conclusions:

• The proposed framework effectively decouples electrode montage from feature extraction, allowing pre-trained models to be used without retraining.
• Enables practical, resource-efficient BCIs that operate without time-locked synchronization or montage-specific calibration.
• Lays the groundwork for bridging the gap between laboratory BCI research and real-world field applications.