EEG Foundation Models Overview

• EEG Foundation Models are large-scale, self-supervised encoders pre-trained on vast, heterogeneous EEG data to generate versatile, generalizable neural features.
• They employ transformer backbones, masking, quantization, and graph neural integration to decode complex spatio-temporal and spectral brain signals.
• EEG-FMs enable efficient few-shot learning and cross-domain generalization, setting new benchmarks in neurodiagnostics, BCI, and cognitive neuroscience.

Electroencephalography Foundation Models (EEG-FMs) are large-scale neural architectures pre-trained with self-supervised objectives on heterogeneous, unlabeled EEG corpora to extract transferable neural representations across a wide spectrum of brain-signal analysis tasks. These models leverage transformer or hybrid backbones, advanced masking and quantization strategies, and inductive architectural biases to surpass traditional, task-specific pipelines in both performance and data efficiency. EEG-FMs have catalyzed progress in BCI, clinical neurodiagnostics, and cross-modal neuroscientific applications, prompting the establishment of new benchmarking standards and a proliferation of foundation-model toolkits for electrical brain signal analysis.

1. Concept and Motivation

EEG-FMs are defined as large-scale, typically self-supervised encoders pretrained on vast EEG datasets to learn generalizable, task-agnostic representations capable of supporting diverse downstream tasks including classification, regression, decoding, and generation (Kuruppu et al., 15 Jul 2025, Xiong et al., 25 Aug 2025, Liu et al., 25 Jan 2026). The rationale for EEG-FMs emerges from the limitations of conventional supervised pipelines, which are hampered by costly expert annotation, low SNR, significant inter-subject variability, and fragmentation across device, protocol, and task domains.

Key motivations:

• Transferability: Pretraining on massive, heterogeneous EEG (and sometimes multimodal) data produces generic representations, reducing the need for per-task model retraining and large annotated datasets (Kuruppu et al., 15 Jul 2025, Shen et al., 12 Feb 2026).
• Data efficiency: EEG-FMs enable effective few-shot and zero-shot learning, critical in domains where labeled data is rare or expensive (Xiong et al., 25 Aug 2025, Li et al., 21 Aug 2025).
• Cross-domain generalization: Foundation models bridge modalities (EEG, audio, text, vision) and support downstream tasks ranging from native decoding to cross-modal generation (Li et al., 21 Aug 2025).
• Unified interface: Foundation models offer a plug-and-play backbone for BCI, clinical, and cognitive neuroscience applications, driving standardization and ecosystem development (Shen et al., 12 Feb 2026, Liu et al., 25 Jan 2026).

2. Architectural Principles and Model Variants

EEG-FMs incorporate design elements tailored to the unique spatio-temporal, spectral, and topological structure of EEG signals. Key architectural developments include:

• Transformer Backbones: Deep (≥6–24 layer) self-attention networks process patchified EEG sequences, often with learnable tokenization and advanced positional encoding (Kuruppu et al., 15 Jul 2025, Zhou et al., 29 Jun 2025, Turgut et al., 28 Feb 2025).
• Multi-scale and Small-world Designs: Models such as CSBrain (Zhou et al., 29 Jun 2025) alternate cross-scale spatio-temporal tokenization with structured sparse attention, capturing both local and global neural dependencies and matching the known mesoscale network organization of the brain.
• Domain Decoupling: Uni-NTFM physically separates time and frequency stream encoders before fusion and incorporates a topological spatial embedding to maintain anatomical relationships, combined with a Mixture-of-Experts Transformer for scalable functional specialization (Chen et al., 29 Sep 2025).
• Graph Neural Integration: GEFM employs a graph neural front-end to propagate information between physically connected electrodes prior to temporal encoding, directly embedding spatial topological priors (Wang et al., 2024).
• Residual Quantization and Hierarchical Coding: BrainRVQ and CodeBrain introduce dual- or multi-domain vector quantizers in the pre-tokenization stage to disentangle time and frequency patterns, further employing hierarchical autoregression or multi-scale architectures for robust discrete representation learning (Cui et al., 18 Feb 2026, Ma et al., 10 Jun 2025).
• Progressive MoE and Spatial Pooling: NeurIPT implements progressive mixture-of-experts subnetworks and intra/inter-lobe pooling, leveraging 3D electrode geometry for montage-invariant representations and regionally interpretable summaries (Fang et al., 18 Oct 2025).
• Generalist TSFMs: Models such as Mantis, trained on cross-domain time series data from heterogeneous sensor types, achieve competitive EEG performance even without EEG-specific inductive bias when followed by task-specific fine-tuning (Gnassounou et al., 31 Oct 2025).

Summary Table: Model Architectural Innovations in Prominent EEG-FMs

Model	Key Feature(s)	Spatial/Bio Priors
CSBrain	Cross-scale tokenization, SSA	Anatomical region partition
Uni-NTFM	Decoupled time/freq, MoE	Channel/region embedding
GEFM	Graph Neural Net front-end	Spherical head electrode graph
BrainRVQ	Dual-domain RVQ, AR pretrain	Importance-aware masking
CodeBrain	TF-Dual Tokenizer, EEGSSM	Small-world (SGConv+SWA)
NeurIPT	AAMP, PMoE, 3D PE, IILP	3D electrode coordinates
Mantis	Per-channel Transformer	None, general TSFM

These innovations are empirically shown to enhance representation quality, robustness to montage/task variation, and biological plausibility in feature learning (Zhou et al., 29 Jun 2025, Cui et al., 18 Feb 2026, Chen et al., 29 Sep 2025, Fang et al., 18 Oct 2025, Ma et al., 10 Jun 2025).

3. Pretraining Strategies and Objective Functions

Self-supervised learning (SSL) is the dominant paradigm in EEG-FMs. The principal pretraining objectives include:

• Masked Signal Modeling (MSM) and Masked Autoencoding: Random or importance-guided masking of time or channels, followed by reconstruction; dominant strategy among AE-based models (Kuruppu et al., 15 Jul 2025, Zhou et al., 29 Jun 2025, Xiong et al., 25 Aug 2025).
• Contrastive Learning: InfoNCE or CPC variants align augmentations (across time, frequency, channels, or modalities) to enforce invariant representations (Kuruppu et al., 15 Jul 2025, Gnassounou et al., 31 Oct 2025).
• Vector Quantization and Codebooks: Dual-domain tokenizers with discrete latent codes (e.g., VQ-VAE, DD-RVQ) for compact, interpretable representations. Cross-entropy over code indices supplements or replaces reconstruction loss (Cui et al., 18 Feb 2026, Ma et al., 10 Jun 2025).
• Autoregressive Modeling: Hierarchical decoding of quantized token sequences in a coarse-to-fine manner, as in BrainRVQ (Cui et al., 18 Feb 2026).
• Hybrid and Cross-modal Objectives: Joint time and frequency domain reconstruction, contrastive-alignment to CLIP or wav2vec embeddings, or generative decoding for cross-modal tasks (Li et al., 21 Aug 2025, Chen et al., 29 Sep 2025).

Curriculum or physiologically-informed masking (e.g., amplitude-aware masking in NeurIPT; importance-guided strategy in BrainRVQ) is increasingly adopted to preferentially force reconstruction of semantically rich neural events, rather than background or artifact-dominated segments (Fang et al., 18 Oct 2025, Cui et al., 18 Feb 2026).

Pretraining datasets are typically composed of orders of magnitude more data than classic supervised regimes—thousands to tens of thousands of hours, spanning multiple public EEG repositories, BCI benchmarks, and sometimes iEEG/fNIRS (Kuruppu et al., 15 Jul 2025, Shen et al., 12 Feb 2026, Chen et al., 29 Sep 2025).

4. Evaluation, Benchmarking, and Empirical Findings

Benchmarking EEG-FMs is standardized by the emergence of frameworks such as EEG-FM-Bench (Xiong et al., 25 Aug 2025) and Brain4FMs (Shen et al., 12 Feb 2026), which harmonize preprocessing pipelines, task definitions, and reporting protocols. Benchmarks span canonical BCI, clinical, and cognitive paradigms:

• Motor Imagery
• Emotion Recognition
• Sleep Staging
• Seizure/Event Detection
• Abnormality/Pathology Classification
• Mental Workload/Stress
• Visual/Auditory Decoding
• Cross-modal Retrieval or Generation

Metrics include balanced accuracy, weighted F1, Cohen’s κ, AUROC/AUC-PR, and task-specific scores (regression r or RMSE). Subject-independent splits and consistent artifact handling are standard (Xiong et al., 25 Aug 2025, Shen et al., 12 Feb 2026, Liu et al., 25 Jan 2026).

Notable findings:

• Generalization Gap: Linear probes of frozen backbones often underperform full fine-tuning, with notable exceptions (e.g., SSVEP paradigms) (Liu et al., 25 Jan 2026, Xiong et al., 25 Aug 2025).
• Supervised Baseline Competitiveness: Lightweight specialist models remain top-3 on many tasks; foundation models match or slightly surpass them after domain-adapted fine-tuning (Liu et al., 25 Jan 2026).
• Architectural Impact: Models integrating fine-grained spatio-temporal attention, cross-scale structure, and neurophysiological priors (e.g., CBraMod, CSBrain, NeurIPT, Uni-NTFM) yield state-of-the-art results, especially under domain shift or label scarcity (Xiong et al., 25 Aug 2025, Zhou et al., 29 Jun 2025, Fang et al., 18 Oct 2025, Chen et al., 29 Sep 2025).
• Pretraining Data and Scale: Beyond a point, scale alone does not guarantee superior performance; the impact of diversity and domain mismatch outweighs raw hours (Kuruppu et al., 15 Jul 2025, Liu et al., 25 Jan 2026).

In certain cases, cross-domain TSFMs, even when pretrained on non-neural or synthetic data, achieve competitive or superior performance to EEG-specific FMs, especially after full fine-tuning (Gnassounou et al., 31 Oct 2025).

5. Biological and Neuroscientific Interpretability

EEG-FMs increasingly provide interpretable, biologically aligned representations:

• Channel/Region Attention Maps: Saliency analyses reveal meaningful localization (e.g., motor cortex in MI, prefrontal in emotion, occipital in alpha-power) (Xiong et al., 25 Aug 2025, Ma et al., 10 Jun 2025, Zhou et al., 29 Jun 2025).
• Codebook Decoding: Temporal/frequency tokens often map to canonical neural events (e.g., delta in N3, alpha in eyes-closed) and are highly class-specific (Cui et al., 18 Feb 2026, Ma et al., 10 Jun 2025).
• Latent Cluster Analysis: t-SNE and embedding diagnostics show that foundation representations produce more compact, task-separable feature spaces (Xiong et al., 25 Aug 2025, Chen et al., 29 Sep 2025).
• Prototype-guided adaptation: Lightweight, structure-aware adapters (e.g., SCOPE) retain pretrained manifold structure and enhance label efficiency by modulating layers according to learned class prototypes (Ma et al., 19 Feb 2026).

Such interpretability is critical for clinical translation, neurocognitive validation, and regulatory acceptance.

6. Open Problems, Limitations, and Future Directions

Current challenges include:

• Label Scarcity: Full fine-tuning with limited labeled subjects leads to collapse or overfitting; structured prototype-based adapters are promising but not yet universal solutions (Ma et al., 19 Feb 2026).
• Pretraining/Objective Limitations: Masked reconstruction of raw signal may encourage memorization of noise. Objectives encoding neuro-semantic structure (contrastive, cross-modal, event-aware masking) are more robust (Liu et al., 25 Jan 2026, Shen et al., 12 Feb 2026).
• Scaling Laws: No clear trend that larger models produce better downstream transfer under current data and architecture regimes (Liu et al., 25 Jan 2026).
• Cross-Modality: Most models are unimodal; cross-modal FMs (EEG–text, EEG–vision, EEG–audio) are under active development with reported performance in open-vocabulary retrieval and generation (Li et al., 21 Aug 2025).
• Standardization and Benchmarks: Continued harmonization of metrics, splits, and datasets is required for reproducibility and fair comparison (Xiong et al., 25 Aug 2025, Shen et al., 12 Feb 2026).
• Deployment and Personalization: Need for parameter-efficient adaptation (adapters, LoRA), online/federated updates, and robustness across devices and populations.

Recommended directions:

• Multimodal, multi-paradigm pretraining and harmonized corpora (Shen et al., 12 Feb 2026)
• Neurophysiology-informed masking/diverse objectives (Cui et al., 18 Feb 2026, Fang et al., 18 Oct 2025)
• Topological/channel-agnostic spatial embeddings and GNN-like modules (Wang et al., 2024, Chen et al., 29 Sep 2025, Fang et al., 18 Oct 2025)
• Instruction/prompt-tuning for clinical interpretability and closed-loop BCI control (Liu et al., 25 Jan 2026, Li et al., 21 Aug 2025)
• Curated, large-scale public EEG/iEEG datasets and transparent software toolkits (Shen et al., 12 Feb 2026, Xiong et al., 25 Aug 2025)

EEG Foundation Models represent an emergent standard for scalable, transferable, and interpretable EEG analysis, but their full potential and ecosystem maturity will require advances in neurophysiological bias integration, cross-modal capability, calibration-free usability, and rigorous benchmarking.