EEG Foundation Models: A Critical Review of Current Progress and Future Directions

Premise. Patterns of electrical brain activity recorded via electroencephalography (EEG) offer immense value for scientific and clinical investigations. The inability of supervised EEG encoders to learn robust EEG patterns and their over-reliance on expensive signal annotations have sparked a transition towards general-purpose self-supervised EEG encoders, i.e., EEG foundation models (EEG-FMs), for robust and scalable EEG feature extraction. However, the real-world readiness of early EEG-FMs and the rubrics for long-term research progress remain unclear. Objective. In this work, we conduct a review of ten early EEG-FMs to capture common trends and identify key directions for future development of EEG-FMs. Methods. We comparatively analyze each EEG-FM using three fundamental pillars of foundation modeling, namely the representation of input data, self-supervised modeling, and the evaluation strategy. Based on this analysis, we present a critical synthesis of EEG-FM methodology, empirical findings, and outstanding research gaps. Results. We find that most EEG-FMs adopt a sequence-based modeling scheme that relies on transformer-based backbones and the reconstruction of masked temporal EEG sequences for self-supervision. However, model evaluations remain heterogeneous and largely limited, making it challenging to assess their practical off-the-shelf utility. In addition to adopting standardized and realistic evaluations, future work should demonstrate more substantial scaling effects and make principled and trustworthy choices throughout the EEG representation learning pipeline. Significance. Our review indicates that the development of benchmarks, software tools, technical methodologies, and applications in collaboration with domain experts may advance the translational utility and real-world adoption of EEG-FMs.

💡 Research Summary

This paper presents a critical review of the emerging field of EEG foundation models (EEG‑FMs), which aim to learn generalizable, task‑agnostic representations of electroencephalography (EEG) data using self‑supervised learning (SSL). The authors begin by outlining the limitations of traditional EEG analysis: expert visual inspection remains the gold standard, handcrafted features are limited, and supervised deep‑learning approaches suffer from data scarcity, over‑fitting, and poor transferability across recording sites, hardware, and subject populations. To address these challenges, the community has begun to develop EEG‑FMs—large neural encoders pretrained on massive unlabeled EEG corpora that can be fine‑tuned with only a few labeled examples.

The review focuses on ten early EEG‑FMs published between January 2021 and September 2024. Each model is examined through three “pillars” of foundation modeling: (1) representation of input data, (2) self‑supervised pre‑text task design, and (3) evaluation strategy. The authors systematically compare the models on these dimensions and synthesize the findings.

Input representation – Most models ingest raw multichannel time‑series directly, segmenting recordings into fixed‑length windows (typically 2–10 seconds) and tokenizing them for transformer encoders. A few works transform the data into alternative views such as short‑time Fourier spectra, wavelet coefficients, or spatial covariance matrices to expose frequency or connectivity information. However, the impact of these alternative representations on downstream performance is not consistently quantified, leaving an open question about the optimal preprocessing pipeline for EEG‑FMs.

Self‑supervised learning – The dominant pre‑text task across the surveyed models is masked sequence reconstruction (a generative SSL approach). Models randomly mask a proportion of the temporal‑spatial tokens and train a decoder to reconstruct the missing values, using losses such as L2, L1, or spectral‑domain errors. Variations exist in mask size (continuous blocks vs. random points), mask ratio, and reconstruction targets, but no systematic ablation studies are provided. Contrastive SSL (e.g., instance discrimination, temporal contrast) appears in only a minority of papers, and multimodal or predictive pre‑texts are largely unexplored. Consequently, the relationship between SSL design choices and the semantic content of the learned embeddings remains under‑investigated.

Evaluation strategy – This is identified as the most critical weakness. Each paper evaluates its model on a different downstream task (e.g., sleep stage classification, seizure detection, cognitive state decoding) and on a distinct public dataset (TUH, SEED, BCI‑IV, proprietary collections). Metrics are limited to accuracy, F1‑score, or AUC, with little attention to robustness (noise, channel dropout, domain shift) or interpretability (mapping latent dimensions to neurophysiological phenomena). Few studies report large‑scale experiments that examine scaling laws (performance vs. model size or data volume), and most fine‑tuning experiments involve modest parameter updates on small labeled subsets.

From this analysis the authors derive several key insights and recommendations:

1. Standardized input pipelines – Incorporate domain knowledge (frequency bands, spatial connectivity) into a reproducible preprocessing framework, possibly offering multiple canonical views (raw, spectral, covariance) for model consumption.
2. Diversify SSL objectives – Move beyond masked reconstruction to include contrastive learning, temporal prediction, multimodal alignment (EEG + video/audio), and hierarchical objectives that capture both short‑term dynamics and long‑term structure.
3. Unified benchmark suite – Establish a community‑wide benchmark comprising multiple public EEG datasets, a set of heterogeneous downstream tasks, and robustness tests (e.g., synthetic noise, channel removal, cross‑site validation). Include interpretability assessments such as saliency maps or alignment with known EEG biomarkers.
4. Systematic scaling studies – Conduct controlled experiments varying model depth/width and pretraining data volume to identify scaling laws analogous to those observed in vision and language foundation models. This will clarify whether larger EEG‑FMs yield diminishing returns or unlock new capabilities.
5. Open‑source tooling and APIs – Release standardized training scripts, model checkpoints, and downstream adapters to lower the barrier for clinicians and neuroscientists, fostering collaboration and accelerating translational adoption.

In conclusion, while early EEG‑FMs demonstrate the feasibility of self‑supervised, transformer‑based encoders for EEG, the field is hampered by heterogeneous evaluation practices, limited exploration of SSL design space, and insufficient evidence of scaling benefits. By adopting standardized data representations, richer self‑supervised objectives, comprehensive benchmarks, and open tooling, future EEG‑FMs can progress from research prototypes to reliable, off‑the‑shelf components for clinical decision support, brain‑computer interfaces, and neuroscience research.