GEFM

Two electrodes a centimeter apart on the scalp pick up overlapping activity from the same cortical generators; two on opposite hemispheres usually do not. An EEG model that treats each channel as an independent time series has to infer that geometry from data alone, and the EEG foundation models GEFM positions itself against — BENDR and Neuro-GPT among them — model channels primarily as parallel temporal streams. GEFM supplies the geometry directly, representing the recording montage as a weighted graph whose edges encode how far apart two electrodes sit on the head.

GEFM was developed by Limin Wang, Toyotaro Suzumura, and Hiroki Kanezashi at the University of Tokyo, posted to arXiv in November 2024 and published at IEEE EMBC 2025. It builds on BENDR, a wav2vec-style contrastive EEG model in which raw multichannel signal passes through a six-layer convolutional encoder, part of the convolved features is masked, and a Transformer encoder reconstructs them under a contrastive loss. GEFM inserts a two-layer graph neural network in front of that convolutional encoder, so every representation the rest of the stack sees has already been mixed across electrodes.

The placement matters more than the module. Because the GNN sits inside the pretraining path, it is trained jointly with the encoder and Transformer on unlabeled clinical EEG and then carried into downstream tasks as part of the checkpoint, rather than fitted per cohort as a downstream adapter.

#Key Features

• Graph ahead of the encoder: A two-layer GNN processes the electrode graph before the convolutional encoder, so inter-channel structure enters the representation at the earliest point rather than after channels have already been mixed.
• Distance-based edge weights: The graph is fully connected over 19 nodes, each edge weighted by the reciprocal geodesic distance between two electrodes on a spherical scalp model, encoding the assumption that nearby electrodes interact more strongly.
• Graph module inside pretraining: The GNN is pretrained under the masked contrastive objective and then fine-tuned, so what transfers is a single checkpoint rather than a per-dataset adaptation recipe.
• Sequence length adjuster: A learned linear layer maps a downstream recording to the pretraining sequence length before the graph, letting one checkpoint serve datasets recorded with different trial durations.
• Architecture-agnostic graph slot: GCN, GAT, and GraphSAGE all drop into the same position and are evaluated head to head.

#Technical Details

Preprocessing follows BENDR: 19 channels of the 10/20 montage, resampled to 256 Hz. Pretraining uses version 2 of the Temple University Hospital EEG Corpus, clinical recordings from over 10,000 individuals, downsampled to one-tenth of its size to keep experimentation tractable, with 60-second sequences (20 seconds for the P300 pretraining run). Three downstream datasets, all disjoint from the pretraining corpus, provide the evaluation: MMI motor imagery (105 subjects, scored by accuracy), P300 target detection (9 subjects, AUROC), and ERN error recognition (26 subjects, AUROC). With a linear length adjuster and the "BENDR" downstream configuration, GCN with edge weights is the only variant to beat the baseline on all three, at 0.849 / 0.616 / 0.538 against 0.646 / 0.577 / 0.522. GraphSAGE scores higher on MMI (0.883) and P300 (0.692) but falls to 0.501 on ERN. The linear adjuster outperforms padding throughout, and the "BENDR" configuration, which retains the pretrained Transformer, beats the encoder-only "Linear" configuration in 21 of 24 GNN runs. The implementation uses PyTorch 2.3.1 with PyTorch Geometric 2.5.3; the paper reports no parameter count.

#Applications

GEFM targets brain-computer interface and clinical EEG settings where labeled trials are scarce but unlabeled clinical recordings are plentiful: motor-imagery decoding for assistive control, P300 spellers, and error-related potential detection for catching mistyped speller output. Because the electrode graph is derived from standard 10/20 coordinates rather than learned per study, any lab recording a conventional montage can use the pretrained representation, and the length adjuster absorbs differences in trial duration between the pretraining corpus and the target protocol.

#Impact

GEFM's contribution is a placement argument: inter-channel structure learned during pretraining and carried in the checkpoint, rather than added afterwards on the downstream cohort. The supporting evidence is conference-scale and uneven — only GCN with edge weights improves on all three tasks, several variants in the encoder-only configuration sit near chance at roughly 0.50, and the downstream datasets are small. No code or model weights have been released, so the work is currently reproducible only from the paper's description. The recipe itself is portable: the authors note that the same graph slot can precede the convolutional encoder of other EEG foundation models built the same way. Catalog peers approach electrode geometry from the other direction — LUNA and MMM normalize arbitrary montages into a shared latent space, whereas GEFM fixes the montage and turns inter-electrode distance into an explicit relational prior.