A roadmap for brain-signal foundation models, written for a field with better data than ours

What the work claims

This is a perspective paper, explicitly didactic in intent: it defines what a foundation model is, surveys the handful of MEG-specific and MEG-inclusive models that exist, lays out four construction strategies, lists the ingredients a serious attempt needs, and closes with benchmarking and governance requirements. Its factual spine is verifiable. MEG records the magnetic fields produced by postsynaptic cortical currents; unlike EEG, those fields are barely distorted by skull and scalp, so MEG couples millisecond timing to unusually faithful spatial sensitivity¹.

The existing landscape is small enough to name in full. MEG-GPT, pretrained on the Cam-CAN dataset of about 612 subjects in a GPT-style autoregressive transformer over region-level time series, is called the first explicit general-purpose MEG foundation model. Headache-MEG-FM, trained on about 416 participants, targets migraine classification and is described as the first clinically oriented one. BrainOmni is the prominent multi-modal entry, pretraining jointly over EEG and MEG and reporting that joint pretraining outperforms single-modality models on both. Several other models, including GPT2MEG and Brain-OF, round out the list. The authors' own verdict on the field's maturity is blunt: pretraining corpora remain modest, benchmarking is only beginning, and core design choices are unsettled¹.

The paper's most quoted-worthy claim is a warning dressed as a roadmap item: foundation models trained on brain signals require data-sharing practices broader than anything task-specific neuroimaging has used, and that expansion collides with the fact that neural data are body-derived, historically collected under tightly governed protocols, and now subject to large-scale repurposing and open-ended downstream use, an argument the authors attribute to Hanley et al. (2026). MEG data are scarce, expensive to acquire, geographically concentrated, and governed by heterogeneous access regimes. Their remedy is a concrete list: consent-compatibility review before aggregation, dataset-level governance documentation, transparent rules for redistribution and for derived models, withdrawal mechanisms where feasible, and tiered-access frameworks balancing reuse against privacy, participant dignity, and equitable benefit sharing¹.

How it works

The paper organizes all routes to a MEG foundation model into four strategies, and the taxonomy is useful because each strategy carries a different dependency on data access. Strategy one is MEG-native pretraining, which learns MEG-specific structure directly but demands broad, harmonized, reusable corpora that do not yet exist. Strategy two is adapting an existing EEG foundation model to MEG inputs, cheap but risky: the authors warn of negative transfer, where representations optimized for EEG constrain the model toward features that are suboptimal for MEG's richer spatial organization, and insist this route be treated as an empirical hypothesis, not a head start. Strategy three is continued pretraining of an EEG model on unlabeled MEG, a middle path the authors call possibly the most practical near-term route. Strategy four is cross-modal pretraining from the outset, with BrainOmni as the existence proof that joint EEG-MEG training can beat both single-modality models¹.

The benchmarking section is where the paper stops being a tutorial and starts being a referee. A leaderboard built partly on private data cannot serve as a shared standard, and the authors single out the MEG component of the NeuralBench suite for exactly that defect. They require that benchmark data be excluded from pretraining and fine-tuning, ideally verified at the participant and recording-session level, because the same recordings can leak across dataset releases. They flag a reporting trap: linear probing and full fine-tuning measure different things, and reporting only fine-tuned numbers inflates models whose frozen features are not linearly separable. And they deliver the hype correction in one sentence: on the one broad comparison available, from NeuralBench on EEG, foundation models are only marginally ahead of task-specific ones. Two escape hatches are noted: geometry-aware encodings and pretraining across acquisition systems have both been shown to transfer across different MEG scanners, so vendor lock-in of the dataset is not inevitable, only default¹.

Where a skeptic should push

The most load-bearing assumption in the whole roadmap is that MEG foundation models will earn their keep over strong task-specific baselines, and the paper's own evidence says that case is not yet made. A marginal edge on EEG, a sibling modality with far more data, is a weak prior for MEG, which has less. Every claim about reusable spatiotemporal representations is currently supported by a literature the authors themselves describe as a small number of models, typically each pretrained on a single dataset or paradigm, which limits exactly the breadth of representation the foundation-model thesis depends on. There is a structural conflict the authors surface honestly: MEG data are so scarce that most available data will end up in the pretraining corpus, which leaves evaluation confined to low-data regimes and makes genuine held-out benchmarking hard to stage at all. A skeptic should also ask who performs the governance work the paper demands. Consent-compatibility review and derived-model redistribution rules are labor with no natural funder, and a roadmap that assigns them to the community in general may be assigning them to no one in particular.

Port the consent machinery before corpora form

For platform access, the paper is a preview of the fight organoid intelligence will have over its raw material, fought by a field with better data and more institutional memory. MEG at least has public datasets, curated lists, and a decade of sharing culture to defend. Organoid electrophysiology has nothing like it: large-scale multi-electrode array recordings of living neural tissue are scarcer than MEG by orders of magnitude, concentrated in a handful of vendor platforms and flagship labs, and expensive in a nastier currency, because every recording consumes tissue with a donor attached. When data are this scarce, aggregation is destiny. Whoever assembles the first thousand-subject neural-tissue corpus, whether a cloud platform, a device vendor, or a consortium, becomes the access gatekeeper for the entire downstream field, and the paper's description of MEG's heterogeneous access regimes will read like an elegy for the open alternative.

The ethics port is where the roadmap's machinery becomes urgent rather than prudent. Human MEG participants consented to neuroimaging studies; the paper already treats large-scale repurposing of that data as a new normative context requiring fresh review. Neural organoid donors are one more step removed: they consented to a research study or a clinical procedure, not to immortal cell lines, not to pretraining corpora, and not to commercial derived models that can emit inferences about genotypes and disease risk. The consent-metastasis problem, where a donor's biological signature outlives every use they were told about, is MEG's problem compounded by immortality and compounded again by commerce. The paper's remedies map directly: consent-compatibility review at aggregation time, documented rules for redistribution and derived models, withdrawal mechanisms where feasible, and tiered access. Applied to organoid data, tiered access is not a privacy nicety but the difference between a shared scientific commons and a vendor's proprietary training moat.

The threat is already legible in the mechanism the paper warns about for MEG: derived models. A pretrained model is a compressed, queryable copy of its training data's statistical content. If a commercial organoid-intelligence vendor trains on proprietary neural-tissue recordings without derived-model rules, the donor's biological information leaves the consent envelope not as data but as weights, unrevocable and uninspectable. The opportunity, equally concrete, is timing. Corpora are still small, no incumbent has won, and the benchmarking discipline the paper lays out, participant-level exclusion, matched task-specific baselines, mandatory linear probing, no private-data leaderboards, is cheap to adopt now and ruinously expensive to retrofit after a vendor's leaderboard becomes the reference. Organoid intelligence should borrow two things from this roadmap before its own data war starts: the governance checklist, and the refereeing instinct.

The bottom line

Established: MEG foundation models exist but are few, each typically pretrained on a single modest dataset; cross-scanner and cross-modality transfer have early positive evidence; and the only broad comparison against task-specific models, from EEG, shows foundation models only marginally ahead. Established also: the field's own roadmap says the binding constraints are infrastructural, corpora, benchmarks, and governance, not architectures. Unproven: that MEG foundation models beat strong baselines at all, and that the governance agenda the paper sets will find an owner. What would confirm the field's promise: a fully open, participant-level decontaminated benchmark showing clear gains over matched task-specific models. What would break it: continued reliance on partly private leaderboards, or an aggregation of the scarce data under a proprietary regime before open governance exists. For computing on living neural tissue, the lesson is to install the consent and derived-model machinery now, at corpus-building time, because the alternative is negotiating access with whoever aggregated the data first.