NOMAD-EEG

EEG foundation model

One small model for any electrode montage.

Native Omni-Montage Adaptive Diffusion

NOMAD-EEG reads EEG with any number of electrodes, in any order and of any length. It identifies each electrode only by its 3D position on the scalp. It has 12.0 M parameters and was pre-trained on 147 hours of unlabelled EEG from one corpus. On ten downstream datasets it was compared with seven EEG foundation models, all run under the same protocol.

19 electrodes · Mumtazsame weights, every montage
  • 12.0 Mparameters (REVE: 69 M)
  • 147 hof pre-training EEG, one corpus
  • 1–64electrodes read without retraining
  • 10downstream datasets
  • 7foundation models compared

The model

Electrodes are positions, not slots.

Most EEG foundation models are tied to the channel set and input length they were trained on. A recording with a different cap, fewer electrodes or a different duration has to be remapped, padded or cut to fit.

NOMAD-EEG never sees a channel index. Each electrode enters with its signal and its 3D coordinate from the standard 10–20 and 10–05 systems, and a bipolar derivation is placed at the midpoint of its two electrodes. The network mixes information along time with state-space layers, without patches, and across electrodes with attention, so the output does not depend on the order in which electrodes are listed.

The backbone is a diffusion model. After pre-training, its intermediate activations serve as features for a frozen linear probe or as the starting point for fine-tuning.

Architecture

From raw signal to a 10,240-dimensional embedding

NOMAD-EEG architecture: pre-training by denoising diffusion, the inside of one NOMAD block, and the two downstream readouts. a Pre-training: denoising diffusion on TUEV (147 h) 21 electrodes, 3D positions pₑ clean 30 s window x₀ E × T at 200 Hz noise ε noisy xₖ pₑ, noise level t, corpus c stem NOMAD block × 16 head NOMAD-EEG · 12.0 M parameters predicted ε̂ loss ‖ε̂ − ε‖² target ε xₖ = √ᾱₖ · x₀ + √(1 − ᾱₖ) · ε (cosine schedule) b Inside one NOMAD block (16 identical blocks) LayerNorm + FiLM(t, c) time → one token per 20 ms electrodes φ(pₑ) σ(g) tanh(f) × u next block skip sum S4D along time one kernel spans the whole segment, same weights for every electrode Electrode attention each 20 ms step attends across electrodes, tokens carry φ(pₑ) Gated output u = σ(g) ⊙ tanh(f) residual + skip c Downstream (clean input) gates of blocks 4, 8, 12, 16 frozen probe mean over electrodes 5 windows: mean, std 10,240-d vector linear classifier fine-tuning E × 25 tokens attention pooling class logits

(a) A clean 30 s window is corrupted with Gaussian noise, and the network predicts the noise from the noisy signal, the electrode positions, the noise level and the corpus. A convolutional stem shared by all electrodes gives one 256-dimensional token every 20 ms. (b) In every block, an S4D layer filters each electrode's row along time; attention then mixes each time step across electrodes, which are identified only by Fourier features of their 3D positions. (c) Downstream, the gated activations of four blocks are read either by a frozen linear probe or by a fine-tuned attention head.

How it works

Any number of electrodes, one fixed-size output

Pick a montage from the paper's datasets and follow one recording through the network. The electrode names, token rate and output size are the model's own; the signals are illustrative and the scalp layout is schematic.

    Highlights

    Where NOMAD-EEG leads

    Frozen linear probes unless stated. Every number is from the paper's tables; the full tables, including the tasks where other models are better, are below.

    EEGMMIDB · motor imagery, 4 classes, 64 electrodes0.442bal. acc.

    Best of eight models. Next: DiffEEG 0.418, REVE 0.371.

    Siena · seizure detection0.622macro F1

    Best of eight models. Next: REVE 0.562, EEGDM 0.534.

    Bonn · seizure detection, single channel0.999AUROC

    Best of eight models. REVE 0.998, CBraMod 0.996.

    TUSZ · seizure detection, fine-tuned0.676macro F1

    After fine-tuning, above every model's frozen probe (best: EEGDM 0.642). As a frozen probe NOMAD-EEG ranks third.

    TUEP · epilepsy diagnosis, 200 patients0.640accuracy

    Highest accuracy and macro F1 of the seven models compared, at the patient level.

    TUEV · six EEG event classes0.635Cohen κ

    After fine-tuning, the highest κ and weighted F1 (0.814). The baselines were evaluated as frozen probes.

    BCIAUT-P300 · autism, unseen 8-electrode cap0.718AUROC

    Best of five models. Picks the attended object in 75.5% of blocks after ten runs (EEGDM 61.2%).

    Mumtaz · only 5 of 19 electrodes0.891mean AUROC

    Down from 0.942 with all 19, using the probe trained on the full montage. Electrode order changes no prediction.

    Results

    Full benchmark

    All baselines were run by us from their released weights, each with its native input format. Frozen encoders with a class-balanced logistic-regression probe; brackets give 95% bootstrap confidence intervals. Bold is the best value in each column.

    † the model's pre-training data included this dataset or the corpus it comes fromIndented rows: NOMAD-EEG variants, outside the bold comparison unless stated

    Epilepsy summary and missing electrodes

    Robust when electrodes go missing

    On Mumtaz, a probe trained once on all 19 electrodes was applied unchanged to random subsets. Mean AUROC went from 0.942 with all electrodes to 0.919 with 14, 0.919 with 10 and 0.891 with 5, and no subset fell below 0.877. Listing the electrodes in a different order changed no prediction.

    Radar chart of macro F1 for eight models on five epilepsy tasks: TUSZ, Siena, Bonn, TUEV and TUEP.
    Macro F1 on five epilepsy tasks for all eight models (frozen probes). Red dots mark a model that was pre-trained on that dataset.
    Scalp maps of electrode subsets on Mumtaz with their AUROC, and AUROC and balanced accuracy against the number of electrodes.
    Missing-electrode test on Mumtaz (1,179 test windows). Filled dots are the electrodes given to the model. AUROC stays high; balanced accuracy drops because scores shift relative to the threshold fitted on the full montage.

    Weights

    Pre-trained backbone and code

    The release contains the EMA weights of the NOMAD-EEG backbone used for every result on this page: the diffusion model, 12.0 M parameters, pre-trained on TUEV for about 74 epochs.

    The repository has the model code and scripts for linear probing, fine-tuning and testing on your own data.

    pip install -e .   # after cloning the repository
    
    import nomad_eeg
    model = nomad_eeg.load_pretrained("nomad_eeg_base.pt")  # the downloaded weights
    x = nomad_eeg.preprocess(x_uv, fs=256)               # [N, E, T] microvolts -> 200 Hz
    scaler = nomad_eeg.RobustScaler().fit(x)             # training data only
    feats = nomad_eeg.extract_features(model, scaler.transform(x), channels)

    Citation

    Cite NOMAD-EEG

    @article{helwan2026nomadeeg,
      title   = {{NOMAD-EEG}: Native Omni-Montage Adaptive Diffusion},
      author  = {Helwan, Abdulkader and Henni, Khadidja and Abou-Abbas, Lina},
      journal = {arXiv preprint arXiv:2610.35246},
      year    = {2026}
    }