PaperlayerView exact revision on arXiv

Source-linked AI summary

A Roadmap for MEG Foundation Models

Philipp Thölke, Hamza Abdelhedi, Yorguin Mantilla-Ramos, Fouad Lbakali, Oumayma Gharbi, Catherine Duclos, Annalisa Pascarella, Vanessa Hadid, Oiwi Parker Jones, Karim Jerbi

arXiv:2609.04461v2q-bio.NCcs.AI

TL;DR

MEG foundation models are an emerging route beyond task-specific decoding, but their pretraining corpora and benchmarks remain limited. This perspective maps the field’s design choices and development pathways; current benchmarks still offer narrow task coverage, making reusable data, rigorous evaluation, and responsible sharing central priorities.

Abstract

from arXiv · show

Foundation models are beginning to reshape brain-signal analysis by moving the field beyond task-specific decoding pipelines toward reusable models pretrained on broad neural datasets. Magnetoencephalography (MEG) is a compelling but still underdeveloped target for this shift: it captures human cortical dynamics at millisecond resolution while offering stronger spatial interpretability than EEG, making it especially valuable for source-resolved studies of perception, language, cognition, and clinical brain function. Yet MEG foundation models remain at an early stage, with only a small number of MEG-specific and MEG-inclusive multi-modal models, modest pretraining corpora, and emerging but still limited benchmarks. This perspective lays down the basic concepts needed to understand MEG foundation models and provides a didactic overview of the field's key design choices, including tokenization, sensor- versus source-space representations, sensor-geometry encoding, backbone architectures, self-supervised objectives, and pretraining data. We then offer a roadmap for future development, organized around native MEG pretraining, adaptation of EEG foundation models, transfer from generic time-series models, and multi-modal integration with EEG, fMRI, MRI, behaviour, and stimulus features. We highlight the need for coordinated infrastructure, including diverse and reusable MEG datasets, rigorous evaluation across subjects, sites, tasks, and clinical settings, and responsible data-sharing practices that address consent, privacy, access, and governance.

I. INTRODUCTION

The introduction places MEG foundation models within the shift from task-specific decoding to reusable models pretrained on broad neural data. It explains the foundation-model idea, situates MEG alongside EEG and fMRI efforts, and outlines the field’s emerging directions.

I. INTRODUCTION

  • MEG offers millisecond temporal resolution and a closer link to cortical sources than EEG, supporting study of fast brain dynamics across perception, language, cognition, and clinical neuroscience.The article’s roadmap identifies tokenization, sensor- versus source-space representations, sensor geometry, architectures, objectives, data, and benchmarks as key design choices.Source: §I · p. 1

A. What are foundation models?

  • Foundation models learn reusable representations from broad, heterogeneous data, often through self-supervised prediction, for tasks beyond those specified during training.For MEG, the motivating question is whether pretraining on raw or minimally processed signals yields transferable representations of fast cortical dynamics.Source: A. What are foundation models? · pp. 1–2

B. Background on EEG and fMRI foundation models

  • EEG foundation models are emerging rapidly, while fMRI offers a precedent for self-supervised pretraining on data with well-defined spatial structure.The article treats EEG as an analog pathway for MEG modeling, not as a direct comparison frame.Source: B. Background on EEG and fMRI… · p. 2
  • MEG-specific and MEG-inclusive multi-modal efforts are beginning to emerge, raising the question of whether MEG should be developed natively or within broader brain models.Examples include joint EEG–MEG models, models incorporating fMRI, EEG, and MEG, and frameworks aligning representations across those modalities.Source: B. Background on EEG and fMRI… · p. 2

C. The Case for MEG Foundation Models

MEG foundation models are promising but remain early, with modest pretraining corpora, emerging benchmarks, and unsettled design choices. Their rationale rests on MEG’s temporal resolution and sensitivity to cortical sources.

C. The Case for MEG Foundation Models

  • MEG combines millisecond temporal resolution with spatial sensitivity and a closer link to cortical sources than EEG, supporting analysis of fast, spatially organized brain dynamics.Its magnetic fields are less affected by the skull and scalp than EEG’s electric potentials.Source: C. The Case for MEG Foundation Models · p. 2
  • Broadly pretrained MEG representations could support source-resolved dynamics and adaptation across subjects, stimuli, tasks, recording sites, and clinical populations.The paper highlights naturalistic language and cognition, clinical brain-state inference, and cortical propagation as potential uses.Source: C. The Case for MEG Foundation Models · p. 2
  • MEG foundation models remain at an early stage, with few MEG-specific or MEG-inclusive models, modest pretraining corpora, emerging benchmarks, and unsettled design choices.The field has a clear scientific rationale, but its infrastructure and modeling principles still need to be built.Source: C. The Case for MEG Foundation Models · p. 2

II. PATHWAYS TO MEG FOUNDATION MODELS

The paper presents several routes to MEG foundation models, each trading MEG-specificity against scale and transferability. The choice of route depends on the model’s starting point and the structure it must learn about MEG.

II. PATHWAYS TO MEG FOUNDATION MODELS

  • The pathway figure frames the routes as complementary strategies with different investments in data and compute and different capacity to express MEG-specific structure.Routes inherited from EEG or generic time-series models require less MEG data and compute but constrain MEG-specific expressiveness.Source: §II · pp. 2, 4, Figure 1 · p. 4

A. Different Strategies for MEG foundation models

  • MEG-native pretraining is the most direct and expressive route, but it requires broad, harmonized MEG corpora that remain smaller and more fragmented than EEG resources.Table I distinguishes direct MEG models, clinical MEG models, MEG-inclusive multi-modal models, and precursors.Source: Table I · p. 3, A. Different Strategies for MEG… · p. 4
  • Adapting an EEG model’s input interface enables embedding extraction, a frozen-backbone task head, or end-to-end fine-tuning, but domain differences make successful transfer uncertain.EEG and MEG differ in geometry, spatial sensitivity, source orientation, signal scale, noise, and preprocessing; the paper recommends benchmarking transfer against MEG-native pretraining.Source: A. Different Strategies for MEG… · p. 4
  • Continued pretraining adds self-supervised learning on unlabeled MEG before supervised tasks, potentially combining EEG scale with adaptation to MEG structure.Whether EEG-derived representations help or remain influential should be tested against MEG-native models.Source: A. Different Strategies for MEG… · p. 5
  • Joint EEG–MEG pretraining can learn shared features while retaining modality-specific information; BrainOmni reports that joint pretraining outperforms single-modality models on both modalities.This route includes MEG alongside other modalities from the outset rather than treating it only as a transfer target.Source: A. Different Strategies for MEG… · p. 5
  • Broader multi-modal pretraining can pair MEG with fMRI, MRI, stimulus, behavioural, or language information, but abundant modalities may dominate and leave MEG peripheral.The strategy is suited to questions requiring complementary information.Source: A. Different Strategies for MEG… · p. 5
  • Generic time-series models may offer useful architectures or pretraining recipes, but MEG-specific geometry, channel structure, artifacts, sampling, and neurobiology require adaptation.Audio-pretrained Whisper representations have also been used for MEG-to-speech or MEG-to-text decoding.Source: A. Different Strategies for MEG… · p. 5

B. Sensor- and source-space representations

Sensor- and source-space representations are separate design choices available to MEG models, regardless of their pretraining route. Each offers benefits and introduces distinct constraints.

B. Sensor- and source-space representations

  • Sensor-space models use recorded channels directly, avoiding commitment to an inverse solution and supporting pretraining on raw or minimally processed data.They must handle differences in channel number, position, orientation, type, and missing sensors across systems.Source: B. Sensor- and source-space… · p. 5
  • Source-space models project MEG onto cortical locations, regions, or parcels, potentially improving anatomical interpretability and alignment across MEG systems.This is attractive when the goal is to relate representations to cortical regions, networks, or source-level dynamics.Source: B. Sensor- and source-space… · p. 5
  • Source reconstruction depends on head models, noise covariance, regularization, and each subject’s anatomical MRI, and shared reconstruction choices can contribute to apparent cross-system alignment.The paper motivates robustness tests across inverse solutions and spatial representations.Source: B. Sensor- and source-space… · p. 5

C. Choosing the best strategy

No single strategy suits every MEG foundation-model use case; the choice depends on the scientific question, available data, anatomical interpretability, and evaluation setting. Tokenization must also match the backbone and objective, with spatial and temporal choices shaping how MEG enters the model.

C. Choosing the best strategy

  • Strategy choice depends on the scientific question, available data, desired anatomical interpretability, and evaluation setting.Clinical classification may use continued pretraining and supervised fine-tuning, while cortical-dynamics interpretation may benefit from source-space or hybrid representations.Source: C. Choosing the best strategy · p. 5
  • Cross-modal pretraining supports richer modality pairings but requires complex processing pipelines; adapting EEG models is computationally lighter but may yield lower performance.The paper frames the choice as a trade-off among initialization, representation space, and objective rather than a universally best pathway.Source: C. Choosing the best strategy · p. 6

III. KEY INGREDIENTS OF A MEG FOUNDATION MODEL

  • MEG input representation should suit the backbone and pretraining objective because it shapes transferability, sample efficiency, recoverable structure, and downstream tasks.The roadmap focuses on native-MEG pretraining while also pointing to adaptation from neighbouring modalities such as EEG.Source: §III · p. 6

A. Data input and tokenization

  • Source-space tokenization can aggregate activity into atlas-defined or learned ROIs, while temporal tokenization ranges from fixed patches to continuous encoders.MEG-GPT uses 52 atlas-based ROIs; simple fixed discretization can perform comparably to learned tokenizers across several evaluation settings [27]. Encoding anatomical precision is a potential advantage of source-space representations.Source: A. Data input and tokenization · p. 6

B. Backbone architectures

Transformers dominate current MEG foundation models, but their prevalence does not establish that they are optimal for MEG. Their adaptations address long, heterogeneous inputs, while explicit positional and sensor-geometry encodings help represent token order and recording-system differences.

B. Backbone architectures

  • Transformers dominate MEG foundation-model backbones, with perceiver-style modules compressing large or heterogeneous inputs into fixed latent representations.Examples include MEG-GPT’s autoregressive perceiver, BrainOmni’s Criss-Cross Transformer, and Brain-OF’s DINT attention and sparse mixture-of-experts backbone; BBL provides a convolutional contrast.Source: B. Backbone architectures · p. 6
  • Transformer prevalence does not show that transformers are inherently optimal for MEG; convolutional backbones remain a viable alternative for continuous signals.The paper notes room for alternatives tailored to raw MEG’s temporal and spatial structure.Source: B. Backbone architectures · p. 6

C. Positional encoding and sensor awareness

  • MEG transformers need explicit temporal and spatial position information, with spatial encoding complicated by sensor-versus-source representations and vendor-specific channel layouts and types.Current schemes include learned, sinusoidal, or RoPE temporal encodings, and categorical identity or 3D position, orientation, and sensor-type encodings for space.Source: C. Positional encoding and sensor… · pp. 6–7
  • Cross-vendor transfer favors general spatial encodings, while wearable and reconfigurable OPM arrays may require representations that accommodate changing sensor geometry.Convolutional encoders such as BBL avoid positional encoding but are typically tied to a fixed channel layout.Source: C. Positional encoding and sensor… · pp. 6–7

D. Self-supervised objectives

MEG foundation models use varied self-supervised objectives, but their corpora and evaluation evidence remain limited. Emerging benchmarks improve comparability, though current MEG coverage is narrow and the evidence base is still developing.

D. Self-supervised objectives

  • Self-supervised pretraining lets models learn from unlabeled data by predicting withheld or transformed inputs or aligning representations with paired modalities.Although MEG recordings often carry task or condition labels, this approach supports unified data ingestion and is reported to improve generalization.Source: D. Self-supervised objectives · p. 7
  • MEG foundation models use masked reconstruction, autoregression, cross-modal alignment, and surrogate-feature prediction, while several objective classes still lack MEG instances.The unrepresented classes identified here are latent-space masked reconstruction, spectral reconstruction, and contrastive predictive coding.Source: D. Self-supervised objectives · p. 7

E. Pretraining data

  • Current MEG models typically pretrain on a single dataset or paradigm, limiting representation breadth; corpus quality, composition, and balance may matter alongside scale.Early evidence suggests selecting datasets by standalone decodability can outperform indiscriminate pooling [18].Source: E. Pretraining data · p. 7

IV. ASSESSING THE PERFORMANCE OF MEG FMS

  • Whether large-scale pretraining transfers across subjects, tasks, and sites remains unclear, and recently available standardized benchmarks still provide a thin evidence base.Matched-protocol EEG studies have not reliably found better generalization from larger models [30],.Source: §IV · p. 7

A. Existing benchmarks and evaluations for MEG FMs

  • The PNPL competitions provide standardized speech-decoding benchmarks, but their narrow task focus limits their coverage of foundation-model claims.The 2025 competition uses LibriBrain for speech detection and phoneme classification; the 2026 competition adds word classification and between-subject evaluation using LibriBrain100 [32]–[34].Source: A. Existing benchmarks and evaluations… · pp. 7–8
  • NeuralBench’s MEG extension includes two tasks, and REVE pretrained on EEG alone performs best among tested models on the MEG typing task.The result offers early evidence of cross-modal electrophysiological transfer, but the MEG benchmark remains preliminary; standardized splits and public leaderboards encourage comparability.Source: A. Existing benchmarks and evaluations… · p. 8

B. What standardized MEG-FM benchmarking would need

A useful MEG-FM benchmark must test the transfer claims it is meant to assess, using diverse, open data and comparable evaluation protocols. Scientific value also requires testing whether learned patterns yield reproducible insight.

B. What standardized MEG-FM benchmarking would need

  • A benchmark with fixed conditions cannot test whether self-supervised pretraining yields reusable representations that transfer across conditions.Its data should be open and span the cognitive, clinical, resting/task, and acquisition-system diversity the model claims to generalize over.Source: B. What standardized MEG-FM… · pp. 8–9
  • Benchmark data should be excluded from pre- and post-training, ideally verified at participant and session level, although scarce MEG data constrain large evaluation sets.For individual-level applications, evaluation should also test stability across repeated recordings despite session variation.Source: B. What standardized MEG-FM… · p. 9
  • Comparisons need standardized protocols, both linear probing and full fine-tuning, and matched hand-crafted and task-specific baselines.Fine-tuned scores alone may overstate the quality of frozen features, and the available broad EEG comparison found foundation models only marginally ahead of task-specific models.Source: B. What standardized MEG-FM… · p. 9
  • Scientific value should also be assessed by whether learned patterns generate hypotheses, replicate across independent datasets, and withstand experimental validation.Applying models to concrete neuroscientific questions is presented as essential for translating their representations into reproducible insight.Source: B. What standardized MEG-FM… · p. 9

V. OPEN QUESTIONS AND FUTURE PATHS FOR MEG FMS

The roadmap identifies corpus construction as a central constraint and calls for coordinated decisions about data composition, preprocessing, source reconstruction, and evaluation.

A. Building usable, diverse pretraining corpora

  • An open, aggregated corpus spanning thousands of subjects and hours across sites and vendors is a plausible near-term goal for MEG foundation-model development.Corpus composition may matter as much as raw volume, while MEG is currently limited on both fronts.Source: A. Building usable, diverse… · p. 9
  • Minimal, standardized preprocessing may preserve learnable signal and improve transfer across sites, while source reconstruction trades comparability and anatomical interpretability against forward- and head-model assumptions.The source-space decision is specific to MEG; projecting to cortical space can improve cross-system comparability and anatomical interpretability.Source: A. Building usable, diverse… · p. 9
  • Benchmark datasets must be excluded from pretraining, so corpus and benchmark construction need coordination from the outset.Because MEG data are scarce, most datasets may enter pretraining, leaving evaluation only in low-data regimes.Source: A. Building usable, diverse… · p. 9

B. Pretraining objectives matched to MEG’s structure

MEG pretraining objectives and tokenization remain open design choices: they must account for predictable signal structure, measurement variability, and the scale and geometry of sensor- and source-space data.

B. Pretraining objectives matched to MEG’s structure

  • MEG objectives should suppress acquisition-related variability while preserving meaningful individual differences, and current models leave several self-supervision approaches underexplored.The surveyed objectives include masked modeling, autoregressive prediction, and paired-modality alignment; latent-space prediction, spectral reconstruction, and contrastive predictive coding remain underexplored.Source: B. Pretraining objectives matched to… · p. 9
  • Low single-trial signal-to-noise makes raw-signal prediction costly, while MEG-GPT and GPT2MEG use quantization before autoregressive prediction.Alignment objectives instead use paired streams and contrastive representations of the same underlying dynamics.Source: B. Pretraining objectives matched to… · p. 10
  • EEG results offer a tentative prior, but the objective choice remains unclear without comparisons matched for data, compute, and architecture.Latent-space prediction has reported competitive cross-subject performance at substantially smaller scales than raw-data reconstruction approaches.Source: B. Pretraining objectives matched to… · p. 10

C. Tokenization and continuous-time representations

  • Sensor-space tokenization can use short signal patches, but MEG channel encoding must include sensor orientation and channel type as well as position.A single position may carry a magnetometer and two orthogonal planar gradiometers that measure different quantities in different units.Source: C. Tokenization and continuous-time… · p. 10
  • Raw patches are not established as the best tokenization, and adaptive boundaries remain essentially unexplored in EEG and untried in MEG.Other explored approaches include spectral or convolutional tokenization, while vector-quantized time-frequency tokens have been explored for single-channel EEG.Source: C. Tokenization and continuous-time… · p. 10
  • Source reconstruction expands hundreds of sensors to thousands or tens of thousands of locations, making spatial tokenization necessary; learned spatial groupings remain untried in MEG.Atlas parcellation fixes resolution in advance and inherits atlas assumptions, while learned tokenizers could discover other spatial groupings.Source: C. Tokenization and continuous-time… · p. 10

D. Cross-modal integration with EEG and fMRI

Cross-modal models offer a possible response to MEG’s data limitations, but their gains and scientific meaning remain unestablished. The roadmap calls for matched comparisons and tests of whether learned representations capture cortical dynamics rather than modality-specific measurement statistics.

D. Cross-modal integration with EEG and fMRI

  • BrainOmni, Brain-OF, and Ferrante et al. represent MEG-inclusive multimodal efforts, but end-to-end multimodal training has underperformed the best unimodal model on its own modality [44].Because data availability is asymmetric and existing models lack unimodal-counterpart ablations, gains on MEG and other modalities remain open questions.Source: D. Cross-modal integration with EEG… · p. 10

E. From engineering metrics to neuroscientific insight

  • High decoding accuracy establishes task-relevant information in representations, but it remains unclear whether they capture underlying brain dynamics or measurement and preprocessing statistics.The scientific ambition is to model cognition, not merely modality- and sensor-specific statistics.Source: E. From engineering metrics to… · p. 10
  • Candidate evidence for scientific insight includes cognitive-aligned latent dimensions, recovery of known functional organization, and novel structure verified with conventional analyses.Leaderboard performance does not capture these tests, and interpretability tooling for them is largely unbuilt.Source: E. From engineering metrics to… · pp. 10–11
  • Source-resolved MEG combines spatial fidelity with millisecond timing, while learned spatial aggregation could reveal arrangements beyond atlas boundaries.The best methodology remains open, and MEG scaling-law analyses across compute, model, and data size have yet to be conducted.Source: E. From engineering metrics to… · p. 11

F. Data sharing, consent, and governance

The roadmap treats responsible data sharing and coordinated evaluation as prerequisites for MEG foundation-model development. It also argues that MEG’s temporal and spatial richness offers distinctive scientific and clinical opportunities, while decoding accuracy alone cannot establish meaningful discoveries or clinical value.

F. Data sharing, consent, and governance

  • Responsible MEG sharing calls for consent-compatibility review, governance documentation, transparent redistribution rules, feasible withdrawal, and tiered access.These safeguards address scarce, expensive, geographically concentrated datasets with heterogeneous access regimes and balance reuse with privacy, participant dignity, and equitable benefit sharing.Source: F. Data sharing, consent, and governance · p. 11

VI. CONCLUSION

  • MEG combines millisecond temporal resolution, relatively high spatial precision, and hundreds of channels in typical whole-head systems, yielding information-rich recordings.A brain-to-text study reported stronger decoding performance with MEG than EEG [18].Source: §VI · p. 11
  • Representation space, tokenization, initialization, and objectives shape what models exploit, their transfer across systems, and the interpretability of their representations.Progress also depends on open held-out datasets, standardized evaluation, matched baselines, and coordinating benchmarks with corpus construction and responsible sharing.Source: §VI · p. 11
  • Decoding accuracy alone does not establish discovery of unknown cortical structure or improvement in clinically relevant decisions.A stronger test is independently verifiable latent organization that generates hypotheses, alongside robust prediction of diagnosis, prognosis, or treatment response.Source: §VI · p. 11
  • Consent, privacy, diversity, bias, and governance should be built into data curation and training as MEG infrastructure and benchmarks develop.The paper frames this early development stage as an opportunity to embed these considerations from the outset [45].Source: §VI · p. 11
Loading 2609.04461v2…