Foundation Models in EEG Analysis

• Foundation models in EEG analysis are large pretrained backbones that learn robust, transferable representations from extensive unlabeled data.
• They employ diverse strategies—from masked reconstruction to contrastive pretraining—to tackle both short-window decoding and clinically grounded long-context challenges.
• These models boost transfer learning for tasks like seizure detection, brain-computer interfaces, sleep staging, and multimodal integration, while addressing device heterogeneity.

Searching arXiv for recent EEG foundation-model papers to ground the article with current citations. arxiv_search query: "EEG foundation model review CLEF CSBrain CBraMod LaBraM OTiS Mamba EEG" Foundation models in EEG analysis are pretrained backbones intended to learn reusable representations from large-scale unlabeled EEG, or from broader time-series corpora that transfer effectively to EEG, and then support downstream adaptation across clinical diagnosis, brain-computer interfaces, sleep analysis, event detection, and multimodal decoding. In the current literature, the term covers several distinct regimes: EEG-native self-supervised models trained on corpora such as TUEG or TUSZ; generic time-series foundation models reused as EEG feature extractors; session-scale clinical models aligned with reports and electronic health records; and cross-domain systems that map EEG into text, vision, audio, or broader multimodal semantic spaces (Turgut et al., 28 Feb 2025, Cao et al., 11 May 2026, Li et al., 21 Aug 2025). The field is therefore defined less by a single architecture than by a transfer-learning objective: broad pretraining, task-agnostic representation learning, and reuse across heterogeneous downstream EEG settings.

1. Definitions and problem formulation

In EEG work, a foundation model is typically understood as a large pretrained model that learns reusable representations from unlabeled EEG and transfers them to downstream tasks, especially when labeled data are scarce (Kommineni et al., 26 May 2026). A closely related formulation treats large pretrained models for general time-series analysis as reusable encoders for EEG, with zero-shot, linear-probe, or fine-tuning adaptation regimes (Turgut et al., 28 Feb 2025). A broader survey perspective extends the notion further to pretrained language, vision, audio, and multimodal models whose learned priors are repurposed for EEG decoding, alignment, or generation, yielding a modality-oriented taxonomy that includes native EEG decoding, EEG-text, EEG-vision, EEG-audio, and multimodal EEG frameworks (Li et al., 21 Aug 2025).

A central distinction in the literature is between short-window EEG decoding and clinically grounded session-scale modeling. Many early EEG foundation models inherit assumptions from BCI-style decoding: they operate on seconds-long windows, emphasize signal-intrinsic pretraining objectives, and are evaluated on tasks such as motor imagery, ERP decoding, or seizure/event classification. Clinical EEG interpretation, by contrast, is session-scale and depends on persistence, recurrence, evolution, and contextual information from neurologist reports or EHR. CLEF explicitly argues that these are different problem formulations and redefines the clinical EEG foundation model around long-context session embeddings rather than snippet-level decoding (Cao et al., 11 May 2026).

A second distinction concerns source domain. Some models are EEG-native, pretrained directly on large EEG corpora such as TUEG, TUSZ, or HEEDB (Shi et al., 2024, Panchavati et al., 2 Sep 2025, Cao et al., 11 May 2026). Others test whether generic temporal structure is sufficient for transfer: OTiS, Mantis, and synthetic-pretrained time-series models show that non-EEG or minimally EEG-containing pretraining can still yield strong EEG performance, particularly when the adaptation regime is full fine-tuning (Turgut et al., 28 Feb 2025, Gnassounou et al., 31 Oct 2025). This suggests that “foundation model” in EEG analysis denotes a family of transfer strategies rather than a strictly EEG-only pretraining doctrine.

2. Architectural regimes and inductive biases

Current EEG foundation models differ primarily in how they encode temporal scale, channel topology, montage heterogeneity, and clinical context. Several models treat long-sequence modeling as the dominant bottleneck. Mentality uses a Mamba-based selective state space model, combined with a U-Net-style encoder-decoder and an EEGNet-inspired temporal frontend, to model 10-second windows of 19-channel EEG sampled at 200 Hz. Its architecture couples a 1D convolutional layer with kernel size 100, a linear channel-mixing layer, stacked Mamba blocks, and a reconstruction decoder, with seizure detection performed from a downsampled hidden representation followed by max pooling and two linear layers (Panchavati et al., 2 Sep 2025). FoME likewise emphasizes long temporal context, using 15 patches of 1500 time points each, corresponding to 90 seconds of EEG, together with time-frequency fusion embedding and factorized temporal and lateral attention across channels (Shi et al., 2024).

A separate line of work argues that temporal sequence modeling alone is insufficient because EEG is intrinsically multichannel and topological. GEFM inserts a two-layer GNN before the BENDR encoder so that pretraining representations reflect both temporal dynamics and inter-channel relationships. Its graph is fully connected, with edge weights defined as the reciprocal of geodesic distance between electrodes on a spherical scalp model, and the best configuration is a GCN with edge weights, a linear sequence-length adjuster, and the richer BENDR downstream configuration (Wang et al., 2024). Relatedly, Mentality explicitly notes that state space models like Mamba process each channel independently unless channel interaction is injected, and treats its learned channel-mixing layer only as an initial approximation rather than a principled spatial model (Panchavati et al., 2 Sep 2025).

Another architectural axis concerns spatial heterogeneity across devices and layouts. HEAR addresses this directly with a learnable coordinate-based spatial embedding, a temporal-slice channel attention module, and a spatially guided transformer with relative spatial bias computed from 3D electrode coordinates. Instead of collapsing all datasets to a shared channel subset, it constructs a global channel dictionary spanning 1,132 unique electrodes across more than 150 layouts and represents each electrode by its coordinate, not by a fixed channel index (Chen et al., 14 Oct 2025). NeurIPT adopts a related but distinct strategy: it uses 3D physical coordinates of electrodes, Amplitude-Aware Masked Pretraining, a Progressive Mixture-of-Experts backbone, and Intra-Inter Lobe Pooling during fine-tuning to exploit regional brain structure (Fang et al., 18 Oct 2025). CSBrain introduces yet another bias: cross-scale spatiotemporal tokenization with kernel sizes $\{1,3,5\}$, anatomical region partitioning into Frontal, Central, Parietal, Temporal, and Occipital areas, and Structured Sparse Attention over temporal windows and brain regions, arguing that scale-agnostic dense modeling inherited from NLP and vision is mismatched to neural data (Zhou et al., 29 Jun 2025).

Clinical long-context models define a further regime. CLEF converts full sessions into $19 \times 128 \times 2048$ multitaper spectrogram tensors over 1280 seconds, compresses them with a VQGAN tokenizer into $8 \times 256$ latent grids, and then performs masked token prediction with a ViT-style encoder-decoder. Its distinctive step is Stage II grounding, where the session-level EEG embedding is aligned separately to summarized neurologist reports and structured EHR using symmetric contrastive losses (Cao et al., 11 May 2026). This shifts the architectural focus from generic waveform modeling toward clinically grounded semantics.

3. Pretraining objectives and representation learning

Masked reconstruction remains the dominant pretraining paradigm, but the literature increasingly treats the choice of objective as a first-order design variable rather than an implementation detail. Mentality pretrains by reconstructing raw 10-second EEG windows using a combination of MSE loss and a spectral loss computed in the Fourier domain. The best model achieves MSE $= 0.0063$, whereas removing the spectral loss increases MSE to $0.025$, indicating that purely time-domain reconstruction underfits important frequency content (Panchavati et al., 2 Sep 2025). FoME also uses masked reconstruction, with 40% of patches replaced by a learnable $[MASK]$ embedding and optimized by mean squared error between original and reconstructed signals (Shi et al., 2024). HEAR follows LaBraM-style quantized pretraining, combining neural representation quantization with Fourier spectrum prediction of amplitude and phase (Chen et al., 14 Oct 2025).

Several models refine masked reconstruction rather than abandoning it. NeurIPT proposes Amplitude-Aware Masked Pretraining, where masking is driven by signal amplitude rather than random time intervals, motivated by the claim that BERT-style masking on continuous EEG rewards local interpolation more than robust representation learning. Its pretraining loss is an $\ell_p$-norm reconstruction objective under amplitude-conditioned masking, and in ablation it improves TUEV balanced accuracy from $0.6320$ under random masking to $0.6845$ under amplitude-aware masking (Fang et al., 18 Oct 2025). CLEF replaces pixel-level reconstruction with masked token prediction over a VQGAN codebook and uses label-smoothed cross-entropy on masked positions only, while later augmenting representation learning with report and EHR alignment (Cao et al., 11 May 2026).

An important counterpoint is that reconstruction itself may induce a harmful bias. Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models argues that reconstruction-based pretraining on EEG overemphasizes aperiodic and low-frequency components while underrepresenting oscillatory structure, especially at higher frequencies. Using synthetic EEG, it shows that aperiodic parameters are linearly decodable from embeddings, whereas oscillation frequency is not robustly decodable; on real BCI datasets, subject identity is much more linearly accessible than task label for LaBraM, CBraMod, and CSBrain, while Sleep-EDF is an exception because low-frequency structure is itself task-relevant (Kommineni et al., 26 May 2026). This is one of the clearest controversies in the field: a representation can reconstruct EEG well and still encode nuisance structure more strongly than task-relevant state.

Other pretraining families explicitly avoid direct waveform reconstruction. The HuBERT-style BCI model pretrains on one-minute continuous EEG segments from TUEG by predicting k-means pseudo-labels for masked regions, first using 100 cluster labels derived from multi-taper spectrograms and then 500 labels derived from Stage-1 embeddings. Stage 1 uses 57 labels per segment, Stage 2 uses 618 labels, and the model contains six convolutional layers followed by a 12-layer, 12-head transformer with 96.4M trainable parameters (Ogg et al., 2 Jun 2025). Generic time-series models such as Mantis instead use contrastive pretraining, constructing token representations from base patches, first differences, and patch statistics, then optimizing an InfoNCE loss over augmented views; notably, both real-data and synthetic-only pretraining transfer effectively to EEG classification (Gnassounou et al., 31 Oct 2025). OTiS, treated as a general time-series foundation model, is evaluated in zero-shot, linear-probe, and fine-tuning regimes and demonstrates that broad pretraining outside the EEG domain can still yield meaningful EEG features (Turgut et al., 28 Feb 2025).

A plausible implication is that objective choice and representation target may be at least as important as model scale. The literature contains explicit evidence for reconstruction, token prediction, and contrastive transfer, but it also contains a mechanistic critique showing that not all self-supervised objectives align equally well with EEG physiology (Kommineni et al., 26 May 2026).

4. Pretraining corpora, benchmarks, and transfer protocols

EEG foundation-model studies operate across markedly different data scales and task scopes. EEG-native pretraining often uses TUEG or TUSZ. Mentality uses TUSZ v2.0.1 with a training subset of 579 patients and 2,138 seizure events and a held-out test set of 43 disjoint patients with 469 seizure events, using 19 channels from the standard 10–20 montage, 200 Hz sampling, notch filtering at 60 and 120 Hz, and non-overlapping 10-second windows (Panchavati et al., 2 Sep 2025). FoME aggregates eight scalp and intracranial datasets totaling over 1.7 TB, more than 30,000 recordings, more than 15,000 subjects, and about 26,000 hours, then preprocesses all signals to 250 Hz and 6-second windows (Shi et al., 2024). HEAR constructs a dedicated heterogeneous corpus of 8,782 hours from 20 public datasets, with more than 150 distinct electrode layouts and up to 1,132 electrodes in a unified global dictionary (Chen et al., 14 Oct 2025). CLEF uses HEEDB, a snapshot of 260,604 recordings totaling 2 million hours from 108,341 patients, with 59.3% of sessions linked to neurologist reports, and evaluates on a 234-task benchmark spanning 120 disease phenotype tasks, 99 medication exposure tasks, and 15 EEG feature tasks (Cao et al., 11 May 2026).

Evaluation protocols are similarly heterogeneous. Some works emphasize strict subject or patient disjointness. Mentality’s held-out testing split is patient-disjoint (Panchavati et al., 2 Sep 2025). The HuBERT-style sparse-channel BCI foundation model uses leave-one-participant-out cross-validation for most tasks and leave-one-run-out for participant identification (Ogg et al., 2 Jun 2025). Neuro-GPT evaluates 4-class motor imagery on BCI Competition IV-2a with leave-one-subject-out cross-validation after pretraining on a 20,000-recording, 5656-hour subset of TUH (Cui et al., 2023). Other work emphasizes between-subject splits across many datasets, as in CSBrain’s 11-task, 16-dataset evaluation and NeurIPT’s eight-dataset benchmark suite (Zhou et al., 29 Jun 2025, Fang et al., 18 Oct 2025).

Generic time-series transfer papers use a different source-data logic. OTiS is pretrained on 640,187 time-series samples from 8 domains, dominated by 400,000 ECG samples and 203,340 weather samples, with EEG comprising only 3,367 samples, or 0.53% of pretraining data. The same architecture is also pretrained on EEG only as OTiS$_\text{EEG}$ using 3,367 EEG samples totaling 125 recording hours, allowing a direct comparison between general and domain-specific pretraining (Turgut et al., 28 Feb 2025). Mantis compares a real-data pretraining regime on 1.8 million heterogeneous time-series samples against synthetic-only pretraining on 1 million CauKer-generated samples, then transfers to motor imagery and sleep staging (Gnassounou et al., 31 Oct 2025).

A recurrent methodological issue is that many downstream evaluations are still full fine-tuning on relatively clean benchmarks. A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models argues that this standard protocol obscures whether the pretrained representation itself is useful under realistic biomedical constraints such as limited labeled data, sparse channels, and parameter-efficient adaptation (Kommineni et al., 27 May 2026). That critique has become central to how foundation-model claims are now interpreted in EEG analysis.

5. Empirical evidence across tasks and transfer regimes

The empirical record supports the utility of pretraining, but not the claim that one architecture or one pretraining family is universally dominant. Mentality provides a clear proof-of-concept result for self-supervised pretraining in seizure detection: the pretrained model plus linear classifier reaches AUROC $19 \times 128 \times 2048$0 on a held-out test set, whereas the same model trained from scratch reaches AUROC $19 \times 128 \times 2048$1. Its reconstruction ablation likewise shows the importance of spectral loss, with MSE $19 \times 128 \times 2048$2 with spectral loss versus $19 \times 128 \times 2048$3 without it (Panchavati et al., 2 Sep 2025). Neuro-GPT shows that self-supervised pretraining on TUH improves leave-one-subject-out motor imagery accuracy from $19 \times 128 \times 2048$4 to $19 \times 128 \times 2048$5 in the encoder-only fine-tuning regime, while linear probing improves from $19 \times 128 \times 2048$6 to $19 \times 128 \times 2048$7 (Cui et al., 2023). LCM reports balanced accuracy gains from $19 \times 128 \times 2048$8 to $19 \times 128 \times 2048$9 on BCIC-2A and from $8 \times 256$0 to $8 \times 256$1 on BCIC-2B when moving from non-pretrained to pretrained LCM (Chen et al., 11 Feb 2025).

At larger benchmark scale, several foundation models establish broad transfer utility. CSBrain reports a macro-average of $8 \times 256$2 across 11 tasks and 16 datasets, surpassing CBraMod at $8 \times 256$3, LaBraM at $8 \times 256$4, and BIOT at $8 \times 256$5. The paper attributes much of this gain to stacked $8 \times 256$6 cross-scale tokenization and Structured Sparse Attention, with especially strong improvements on BCIC-IV-2a, BCIC2020-3 imagined speech, MentalArithmetic stress, TUEV, and TUSL (Zhou et al., 29 Jun 2025). NeurIPT reports state-of-the-art results across eight downstream BCI datasets, with particularly large gains on MentalArithmetic and Mumtaz2016, though TUAB remains mixed: balanced accuracy is marginally best at 82.93 versus 82.89 for CBraMod, but AUC-PR and AUROC are lower than CBraMod’s 92.58 and 92.27 (Fang et al., 18 Oct 2025).

Device heterogeneity and unseen-layout transfer have also become explicit empirical targets. HEAR-base reaches balanced accuracy $8 \times 256$7 on BCI-IV-2B, $8 \times 256$8 on LargeMI, $8 \times 256$9 on EEGMMIDB, and 0.9620 on TUAB, outperforming BIOT, BENDR, LaBraM, and EEGPT on the reported benchmarks. Its strongest result is zero-shot transfer across non-overlapping layouts on BCI-IV-2B: HEAR-tiny reaches 0.7190 and 0.7071 accuracy in the $= 0.0063$0 and $= 0.0063$1 settings, versus 0.6292 for EEGPT in both directions (Chen et al., 14 Oct 2025). That result is especially important because it tests generalization over electrode layouts rather than over subjects alone.

Clinically grounded long-context modeling produces the strongest large-scale clinical benchmark gains presently reported. CLEF reconstruction-only reaches overall AUROC $= 0.0063$2, already exceeding all prior EEG foundation baselines, and full CLEF reaches $= 0.0063$3 versus $= 0.0063$4 for NeuroLM. It outperforms prior EEG foundation model baselines on 229 of 234 tasks, and the improvement extends not only to EEG feature tasks but also to disease and medication prediction. External-cohort evaluations likewise show gains on TUAB, TUEP, and HSP (Cao et al., 11 May 2026).

Generic time-series pretraining produces a different but equally consequential empirical picture. OTiS outperforms OTiS$= 0.0063$5 on LEMON age prediction, Epilepsy seizure detection, and TUEV event classification; it exceeds several specialized EEG baselines and BIOT on TUEV, though LaBraM remains stronger when much larger EEG-specific pretraining is available (Turgut et al., 28 Feb 2025). Mantis, pretrained on heterogeneous real-world time series or purely synthetic data, consistently outperforms EEGNet and often CBraMod on sleep staging, while matching or slightly exceeding CBraMod on some motor imagery tasks. On SHU-MI, the synthetic-pretrained variant reaches balanced accuracy $= 0.0063$6 and AUC-PR $= 0.0063$7, exceeding CBraMod on both metrics except AUROC, where CBraMod remains slightly higher (Gnassounou et al., 31 Oct 2025). These studies directly undermine the assumption that EEG-specific pretraining is always necessary.

6. Limitations, controversies, and evaluation gaps

The field’s principal controversy concerns what current pretrained EEG representations actually encode. The most explicit critique is the mechanistic analysis of reconstruction-based EEG foundation models, which shows that embeddings from LaBraM, CBraMod, and CSBrain linearly encode aperiodic parameters much better than oscillatory frequency, and on BCI datasets encode subject identity more strongly than task label. In BCIC-IV 2A, for example, LaBraM yields Cohen’s kappa $= 0.0063$8 for task versus $= 0.0063$9 for subject; analogous gaps appear for CBraMod and CSBrain across BCIC-IV 2A, Kaggle-ERN, and Physionet-MI (Kommineni et al., 26 May 2026). This identifies a concrete failure mode: large-scale reconstruction pretraining can organize embedding space around trait-like subject structure rather than state-like task information.

A second major issue is evaluation protocol. Multi-dimensional evaluation shows that current EEG foundation models are conditionally useful rather than uniformly superior. Under full fine-tuning, they outperform supervised baselines on long-context tasks such as MDD MAL and Sleep EDF, but on short-window BCI-style tasks such as PhysioNet motor imagery, BCIC IV-2A, and Kaggle ERN, compact supervised CNNs like EEGNeX or EEGNet Large remain competitive or better despite orders-of-magnitude fewer parameters. The same framework shows that LoRA often approaches full fine-tuning with only 2–4% trainable parameters, but channel-constrained robustness remains limited (Kommineni et al., 27 May 2026). A plausible implication is that many published gains partly reflect adaptation capacity rather than universally transferable pretrained structure.

Spatial modeling is another unresolved bottleneck. Mentality explicitly states that state space models like Mamba process each channel independently and that its learned channel-mixing layer is only an initial approximation (Panchavati et al., 2 Sep 2025). GEFM shows that adding an explicit graph prior can help across motor imagery, P300, and ERN, but its graph is fixed and geodesic-distance-based rather than learned, and its evaluation scope is limited to three tasks (Wang et al., 2024). HEAR and NeurIPT improve robustness to layout heterogeneity, but both incur additional architectural complexity and do not fully eliminate the need for careful metadata curation or coordinate fidelity (Chen et al., 14 Oct 2025, Fang et al., 18 Oct 2025).

Interpretability and reproducibility remain inconsistent. Several papers acknowledge missing implementation details such as optimizer settings, training schedules, block counts, or hardware, which limits exact replication (Panchavati et al., 2 Sep 2025, Wang et al., 2024, Chen et al., 11 Feb 2025). Cross-domain survey work likewise notes that generative EEG-to-text or EEG-to-image pipelines may produce plausible outputs because language or diffusion priors are strong, not because EEG grounding is necessarily faithful; authenticity and interpretability are therefore open problems, especially in multimodal and generative settings (Li et al., 21 Aug 2025).

7. Clinical, cross-domain, and future directions

A major present direction is the shift from short-window decoding toward clinically grounded representation learning. CLEF argues that clinical EEG interpretation is session-scale, not snippet-scale, and operationalizes that claim with multitaper spectrogram tokens over 1280-second sessions, VQGAN tokenization, and contrastive alignment to summarized reports and structured EHR. Its gains on disease phenotypes, medication exposures, EEG findings, held-out concepts, and external cohorts suggest that clinically grounded long-context learning is now a distinct paradigm within EEG foundation modeling (Cao et al., 11 May 2026). FoME pushes in a different direction by jointly pretraining on scalp and intracranial EEG at very large scale and showing especially strong transfer for forecasting and imputation, implying that latent-dynamics modeling may be as important as classification accuracy for defining a useful EEG backbone (Shi et al., 2024).

A second direction is robust handling of heterogeneity. HEAR frames device heterogeneity as a systems bottleneck for scaling EEG pretraining, and shows that coordinate-based channel representations and layout-aware batching can produce genuine unseen-layout transfer (Chen et al., 14 Oct 2025). NeurIPT extends this by combining coordinate-based spatial encoding, amplitude-aware masking, progressive expert specialization, and region-aware pooling, arguing that EEG foundation models should be tailored to signal physics and physiology rather than imported wholesale from generic transformer design (Fang et al., 18 Oct 2025). GEFM and Mentality further suggest that graph-based or more principled multichannel spatial modeling remains a necessary future step (Wang et al., 2024, Panchavati et al., 2 Sep 2025).

A third direction is cross-domain and multimodal grounding. Survey work organizes this landscape into native EEG decoding, EEG-text, EEG-vision, EEG-audio, and multimodal EEG fusion, with representative systems ranging from CLIP-style EEG-image retrieval and diffusion-based EEG-to-image generation to EEG-to-text generation with BART or GPT-family decoders and multimodal EEG-fMRI-calcium alignment (Li et al., 21 Aug 2025). This suggests that EEG foundation models are no longer confined to signal classification; they are increasingly used as semantic alignment modules, generative conditioners, and representation bridges across modalities.

The literature also points toward several common future requirements. These include broader and more balanced pretraining corpora across neurological conditions and healthy states; stronger handling of heterogeneous montages and missing channels; objectives that explicitly preserve high-frequency oscillatory structure rather than only dominant low-frequency energy; evaluation protocols that test subject invariance, nuisance encoding, sparse channels, and low-label adaptation; and support for variable-length or streaming contexts rather than only fixed windows or fixed sessions (Kommineni et al., 26 May 2026, Kommineni et al., 27 May 2026, Cao et al., 11 May 2026). Taken together, the current evidence suggests that foundation models have become a credible organizing framework for EEG analysis, but only under a nuanced reading: they are strongest when architectural bias, pretraining objective, data heterogeneity, and evaluation protocol are all aligned with the temporal, spectral, spatial, and clinical structure of EEG.