REVE: A Foundation Model for EEG -- Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects

TL;DR

REVE introduces Representations for EEG with Versatile Embeddings, a foundation model designed to generalize across diverse EEG setups. It combines a novel 4D Fourier-based positional encoding with a masked autoencoder objective trained on over 60k hours of EEG data from 92 datasets (25k subjects), enabling flexible handling of arbitrary electrode configurations and sequence lengths. A two-task SSL objective and a spatio-temporal Transformer yield robust, transferable embeddings that significantly improve performance on 10 downstream EEG tasks, including motor imagery, sleep staging, and emotion recognition, with notable gains in linear probing and cross-dataset generalization. The approach demonstrates scalable benefits from larger models, expands transferability to unseen montages, and provides open-source resources to advance standardized EEG research and clinical neuroscience.

Abstract

Foundation models have transformed AI by reducing reliance on task-specific data through large-scale pretraining. While successful in language and vision, their adoption in EEG has lagged due to the heterogeneity of public datasets, which are collected under varying protocols, devices, and electrode configurations. Existing EEG foundation models struggle to generalize across these variations, often restricting pretraining to a single setup, resulting in suboptimal performance, in particular under linear probing. We present REVE (Representation for EEG with Versatile Embeddings), a pretrained model explicitly designed to generalize across diverse EEG signals. REVE introduces a novel 4D positional encoding scheme that enables it to process signals of arbitrary length and electrode arrangement. Using a masked autoencoding objective, we pretrain REVE on over 60,000 hours of EEG data from 92 datasets spanning 25,000 subjects, representing the largest EEG pretraining effort to date. REVE achieves state-of-the-art results on 10 downstream EEG tasks, including motor imagery classification, seizure detection, sleep staging, cognitive load estimation, and emotion recognition. With little to no fine-tuning, it demonstrates strong generalization, and nuanced spatio-temporal modeling. We release code, pretrained weights, and tutorials to support standardized EEG research and accelerate progress in clinical neuroscience.

Figures (1)

• Figure 1: Overview of the REVE pretraining framework. The model processes multi-channel EEG data through a linear Patch Embedding where signals are divided into overlapping temporal patches for each channel and embedded with a linear layer. 4D Spatio-Temporal Position Encoding combines spatial coordinates of electrodes with temporal patch indices, augmented with noise for robust generalization. A Block Masking Strategy masks contiguous regions across spatial and temporal dimensions to simulate realistic disruptions. The transformer encoder processes unmasked embeddings. Updated embeddings are joined with learnable placeholders for the masked tokens, from which raw EEG is reconstructed using the decoder. The Primary Task predicts raw EEG signals directly, while the Secondary Task trains a single global token via attention pooling to summarize the input. Both tasks minimize an $\text{L}_1$ reconstruction loss.