REVE: EEG Representation Model

• REVE is a foundation model for EEG that leverages 4D Fourier positional encoding and masked autoencoder pretraining to improve representation learning.
• It delivers state-of-the-art performance on tasks like event detection, pathology classification, cognitive, and affective assessments.
• REVE robustly generalizes to diverse electrode arrangements and acquisition protocols while enhancing data efficiency in clinical and neuroscience applications.

REVE is the acronym for "Representation for EEG with Versatile Embeddings," a foundation model architecture for electroencephalography (EEG) representation learning that enables robust generalization across heterogeneous acquisition protocols, electrode arrangements, and downstream clinical or neuroscience tasks. Developed to address the limitations of prior EEG foundation models—specifically their poor adaptability to diverse input montages and limited performance under linear probing—REVE introduces a 4D positional encoding scheme and a scalable masked autoencoder pretraining objective. Its large-scale multi-dataset training yields state-of-the-art results across a wide spectrum of EEG tasks, from event detection and pathology classification to cognitive and affective assessments (Ouahidi et al., 24 Oct 2025).

1. Model Architecture and Positional Encoding

REVE is structured as a spatio-temporal Transformer encoder integrated into the encoder-leg of a masked autoencoder. The input signal $\mathbf X\in\mathbb R^{C\times T}$, where $C$ is the number of electrodes and $T$ the temporal dimension, is segmented into overlapping temporal patches of length $w$, yielding a tensor $\mathbf X_p\in\mathbb R^{C\times p\times w}$ with $p$ temporal patches per channel. Each patch is projected into a $D_E$-dimensional embedding space, forming $\mathbf E\in\mathbb R^{C\times p\times D_E}$.

A central innovation is the 4D Fourier positional encoding. For each channel $c$ and patch index $t$, position vectors $C$0 are constructed, where $C$1 are electrode coordinates, $C$2 is added Gaussian noise, and $C$3 is a temporal scaling factor. Discrete frequencies $C$4 per axis give rise to composite frequencies $C$5, leading to a positional encoding feature vector:

$C$6

This Fourier representation is summed with a learnable MLP-based embedding and layer-normalized, providing continuous spatial and temporal positional information that supports arbitrary layout generalization. The encoder employs RMSNorm, GEGLU activations, bidirectional attention, and high-efficiency FlashAttention v2.

2. Pretraining: Masked Autoencoding on Diverse EEG Corpora

The pretraining objective is a masked autoencoding (MAE) setup, where a block-structured mask $C$7 omits $C$8 of channel-temporal patch tokens. The encoder receives only unmasked tokens plus positional encodings. Both the visible tokens and learned mask tokens are passed to a lightweight decoder, which reconstructs the original EEG waveforms for masked patches.

The loss is:

$C$9

with an auxiliary loss from a "summary token" reconstruction, weighted by $T$0.

Training is performed on 61,415 hours of EEG from 92 datasets, covering 24,274 subjects and a wide array of clinical, BCI, and cognitive settings. Montages span from 3-channel to 129-channel systems, and signals are minimally preprocessed and resampled to 200 Hz (Ouahidi et al., 24 Oct 2025).

3. Fine-tuning, Linear Probing, and Adaptation Strategies

REVE supports various adaptation mechanisms post-pretraining:

• Linear probing: The encoder is frozen, and a linear classifier is trained atop pooled patch tokens or a summary token.
• Full fine-tuning: A staged protocol first trains the head with the encoder frozen, then unfreezes and fine-tunes the entire network, optionally using low-rank adapters (LoRA) in attention projections. Optimization uses StableAdamW, large batch sizes, and trapezoidal learning rate schedules.
• Parameter-efficient adaptation: LoRA adapters and model souping (checkpoint averaging) further support efficient transfer to new tasks.

Ablation studies, including those on burst-suppression detection in ICU EEG, demonstrate that full fine-tuning systematically outperforms frozen-backbone approaches (e.g., event-based F1 gain of +0.077 in full fine-tuning over random init, and +0.102 over frozen backbone) (Vasta et al., 18 Jun 2026).

4. Downstream Task Performance and Generalization

REVE achieves state-of-the-art or competitive results across ten standard EEG tasks, reported in balanced accuracy (BA):

Task Category	Dataset/Task	REVE-Base BA	Prior SOTA (Model) BA	Δ (pp)
Motor Imagery	BCI-IV-2a (4c)	63.96	51.38 (CBraMod)	+12.58
Seizure/Event	TUEV (6c)	67.59	66.71 (CBraMod)	+0.88
Sleep Staging	ISRUC (5c)	78.19	78.65 (CBraMod)	–0.46
Abnormal Detection	TUAB Abnormal (2c)	83.15	82.89 (CBraMod)	+0.26
Emotion	FACED (9c)	56.46	55.09 (CBraMod)	+1.37
Mental Stress	MAT (2c)	84.50	79.05 (CBraMod)	+5.45

Averaged across all tasks, REVE-Base exceeds the strongest prior model by +2.5 percentage points in BA. Linear probing on REVE-Large yields a BA of 65.4% versus 50.1% for CBraMod, demonstrating improved linear separability (Ouahidi et al., 24 Oct 2025). REVE also generalizes to unseen montages and input channel counts, sustaining high performance even under extreme channel sparsity.

5. Spatio-Temporal Modeling and Robustness

REVE's 4D Fourier encoding and block-masking strategy enforce robust dependency modeling over both spatial (electrode) and temporal (patch) axes. Masking ensures that the encoder learns to interpolate missing channels and time segments. Ablations confirm that replacing the 4D positional encoding with less expressive alternatives degrades performance by 2–3 percentage points on major tasks (e.g., PhysioNet-MI, Mental Arithmetic).

The model handles arbitrary test-time electrode arrangements and sequence lengths owing to its continuous positional encodings, obviating the need for explicit montage alignment or input reconfiguration.

6. Foundation Model Interpretability

A mechanistic interpretability framework using TopK Sparse Autoencoders (SAE) has been applied to REVE embeddings (Lehn-Schiøler et al., 13 May 2026). Activation vectors from each transformer layer are decomposed into sparse feature dictionaries. These are benchmarked using clinical concept enrichment metrics and a concept steering protocol ("target vs. off-target" area metric) that quantifies disentanglement and selectivity.

Results indicate that in REVE, clinical concepts such as abnormality are encoded but entangled—not selectively steerable—unlike in some specialized models (e.g., SleepFM for age). Spectral decoders map manipulations in embedding space back to EEG amplitude spectra, revealing, for instance, that global delta/theta slow-wave power collapses during abnormal-to-normal interventions, reflecting entangled rather than monosemantic interneuron encodings (Lehn-Schiøler et al., 13 May 2026).

7. Clinical Implications, Data Efficiency, and Limitations

REVE's architecture streamlines EEG representation learning for highly variable clinical, cognitive, and BCI settings:

• Pretraining on large, diverse datasets ensures robust transfer to downstream tasks and supports data efficiency. In burst-suppression detection, the pretrained REVE-Base model achieves an event-based F1 score of $T$1 on ICU EEG and reduces burst-per-minute error by 52.1% over EEGNet, with significant gains in low-data regimes (ΔF1 = +0.723 at 25% labeling) (Vasta et al., 18 Jun 2026).
• Its general-purpose embeddings support rapid BCI calibration, montage-agnostic clinical deployments, and standardized multi-site research with shared feature spaces. Released code, pretrained weights, and coordinate loaders facilitate reproducibility and further development (Ouahidi et al., 24 Oct 2025).
• Observed limitations include entangled concept representations in the embedding space and dependence on sufficient pretraining diversity for optimal transfer. External validation and deployment studies beyond single-center settings are ongoing.

• REVE: A Foundation Model for EEG—Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects (Ouahidi et al., 24 Oct 2025)
• Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders (Lehn-Schiøler et al., 13 May 2026)
• Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU (Vasta et al., 18 Jun 2026)