Montage-Agnostic and Calibration-Free EEG Event Segmentation

ICLR 2026 Conference Submission19566 Authors

Keywords: Electroencephalography, brain-computer interface, montage invariance, foundation models, segmentation

TL;DR: A montage-agnostic layer interpolating learned channel embeddings and a lightweight head turn frozen EEG transformers into cross-dataset, calibration-free segmentors for precise event segmentation, outperforming all baselines on diverse datasets.

Abstract: Accurate segmentation and detection of events in continuous electroencephalography (EEG) signals are critical for advancing brain-computer interfaces (BCIs) and understanding neural dynamics, but generalization across diverse recording montages and datasets remains a major challenge. Traditional methods for EEG analysis have largely relied on manual segmentation or time-locked experiments, which struggle with real-world scenarios that involve irregularly timed stimuli triggering or endogeneous events from the BCI operator. Although progress has been made in segmentation tasks such as sleep staging, those approaches primarily address longer, well-defined segments and do not generalize to short-duration stimuli events, such as events typically used in BCI control. In this work, we introduce a novel montage-agnostic framework with spatial interpolation of learned channel embeddings. This approach enables high temporal resolution segmentation in continuous EEG that operates without training on the target recording montage, requiring no subject-specific calibration. Validated across diverse datasets and different paradigms spanning P300, SSVEP and motor imagery, our model consistently outperforms the original foundation models including BIOT and EEGPT, demonstrating superior cross-dataset and montage-agnostic generalization. By removing montage dependency, our framework lays the groundwork for future real-time applications in neuroscience research, clinical diagnostics, and the development of brain-computer interfaces. The code and pre-processed datasets are available at: https://anonymous.4open.science/r/BN00FFgN2H1mhYhVC10J-4505

Primary Area: applications to neuroscience & cognitive science

Submission Number: 19566