NERVE: Noise-Variability-Robust EEG Foundation Model with Electrode-Brain Interactions

Hyunwoo Seo, Jiwon Kim, Byeongyeon So, MINSEONG KIM, Youngjun Song, Seoyoung Jin, Woon-Hong Yeo, Chiehyeon Lim

Keywords: Foundation model, Electroencephalography, EEG, Self-supervised learning, Pre-training

Abstract: Electroencephalography (EEG) is an indispensable modality for measuring and recording brain electrical activity, with broad applications in brain–computer interfaces (BCI) and healthcare. While early EEG models predominantly adopted supervised learning methods due to the scarcity of large-scale datasets and the heterogeneity across tasks and datasets, the recent success of large foundation models has driven increasing efforts to build EEG foundation models. However, most existing studies focus on handling signals with varying formats while overlooking inherent characteristics of EEG signals during acquisition, including low signal-to-noise ratios (SNR), high variability across samples, and spatial dependencies arising from electrode placement within the acquisition system. To address these challenges, we propose NERVE, a novel noise-variability-robust EEG foundation model with electrode-brain interactions. Specifically, pre-training of NERVE begins with learning a noise-robust neural tokenizer that encodes EEG patches into discrete neural tokens. The tokenizer is trained through denoising temporal–spectral prediction to reconstruct temporal and frequency information of the original signal from noise-augmented inputs. NERVE is further pretrained to predict the neural codes of masked EEG patches, integrated with a variability-robust objective that promotes uniform EEG representations. To incorporate spatial structure in EEG, we propose an electrode-position-aware transformer as the backbone for both the tokenizer and the foundation model. It enables the model to capture spatial dependencies among electrodes and brain regions via attention mechanisms. NERVE demonstrates competitive performance across diverse BCI tasks and improved robustness to noise and variability compared to existing EEG foundation models.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 22991