FoME

An EEG recording made in one lab rarely resembles one made in another. The eight public corpora FoME pretrains on run from 3-channel scalp montages to 192-channel intracranial arrays, cover more than 40 distinct scalp electrode configurations, and sample anywhere from 100 Hz to 32 kHz. Earlier EEG models coped with that spread by projecting every montage onto a fixed topological map or by learning a separate channel encoding per dataset, both of which bind the model to the layouts it was built for. FoME instead makes the attention operation itself adaptive along both of the axes that vary.

Its adaptive temporal-lateral attention scaling (ATLAS) mechanism pairs a temporal encoder, attending across time within a channel, with an adaptive multi-channel encoder attending across channels at each time step. Rescaling attention in both dimensions replaces the fixed spatial prior, so no per-dataset topological rule or custom channel embedding is needed, and a time-frequency fusion embedding folds spectral band power into every patch token.

FoME was developed at the Center for Brain and Brain-Inspired Computing Research at Northwestern Polytechnical University and posted as a preprint in September 2024. It is benchmarked against the EEG foundation models LaBraM, Brant and Neuro-GPT, and differs from them in pooling scalp and intracranial recordings into one pretraining corpus and in covering forecasting and imputation alongside classification.

#Key Features

• Adaptive temporal-lateral attention scaling: attention is scaled along the temporal and channel dimensions independently, letting one set of weights model recordings whose electrode counts and arrangements have nothing in common.
• Time-frequency fusion embedding: each patch token sums a linear embedding of the raw signal, a learnable positional embedding, and a frequency embedding derived from log power across eight bands, from delta (1–4 Hz) through four gamma sub-bands up to 100 Hz.
• Cyclic pretraining over heterogeneous blocks: the corpus is partitioned into 5–10 GB blocks of structurally similar signals, one cycle through all blocks being an epoch; 40% of patches are masked and reconstructed under a squared-error loss.
• One backbone, four task families: swapping the reconstruction head for an MLP classifier or a linear forecasting head adapts the same weights to classification, short- and long-term forecasting, and imputation.
• Scalp and invasive data together: two intracranial corpora recorded at 25 and 32 kHz sit in the pretraining mix beside six scalp datasets, so one checkpoint covers both modalities.

#Technical Details

FoME is a transformer with 12 temporal encoder layers and 4 adaptive multi-channel encoder layers. Two variants differ only in feed-forward width: FoME-Base at 476.3M parameters (feed-forward dimension 3,072) and FoME-Large at 744.8M (7,168), which supplies the reported results. Pretraining draws on TUEG, SEED, SEED-IV, CHB-MIT, Sleep-EDFx, a motor-imagery set, and the invasive MAYO and FNUSA corpora — over 30,000 recordings from more than 15,000 subjects aged 1 to 90, roughly 26,000 hours and 1.7 TB. Signals are notch-filtered, band-passed at 0.5–100.5 Hz, resampled to 250 Hz and cut into 6-second windows. Training ran 1,096k steps over 350 hours on six NVIDIA RTX 4090 GPUs.

On seizure detection FoME records 95.15 F2 on MAYO and 91.80 on FNUSA, and across the four epilepsy subtasks spanning detection and classification it leads LaBraM on F2 in three, by 1.71 percentage points on average. It reaches 58.80% accuracy on three-way SEED emotion classification and 99.03% on SleepEDFx sleep staging. Its widest margins are in the regression tasks: on MAYO imputation at a 40% mask ratio it records 0.1913 MAE against 0.3381 for PatchTST, and it leads every baseline on short-horizon (12 s) and long-horizon (30 s) forecasting from a 90-second context. Ablations on TUEV show the temporal encoder contributes most of the gain.

#Applications

The evaluated tasks map onto workflows that already consume EEG: seizure detection and typing in epilepsy monitoring units, automated sleep staging, affective-state decoding, and brain-computer interfaces. The forecasting head targets pre-ictal early warning, where 12 to 30 seconds ahead is the useful horizon, and the imputation head reconstructs channels lost to electrode movement or equipment fault. Because one checkpoint transfers across montages, groups holding modest labeled cohorts can fine-tune rather than train a network from scratch.

#Impact

FoME's contribution is architectural: scaling attention adaptively across time and channels is enough to absorb montage heterogeneity, removing the per-dataset channel bookkeeping that constrained earlier EEG encoders, and it pushes foundation-model evaluation on EEG beyond classification into forecasting and imputation, where its margins are largest. Practical uptake is limited by what has been released: the repository named in the paper exists under an Apache 2.0 license but holds only that license and a one-line README, and no weights have been published, so the results cannot be reproduced or the checkpoint reused. The preprint remains under review with a single posted version. The authors note further limits: the pretraining mix underrepresents emotion, sleep, and motor-imagery recordings and omits cognitive tasks such as working memory, invasive EEG is underused relative to its signal-to-noise advantage, and the 90-second input window caps the context available to the forecasting head.