Organizations: Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon · Institute of Applied Artificial Intelligence, TÉLUQ University, Montreal, Canada
Purpose. Electroencephalography (EEG) montages, channel counts and recording lengths vary widely, but most EEG foundation models are bound to the channel set and input length they were trained on. We aimed to build a generative EEG foundation model independent of the number, identity and order of electrodes and of the input length. Methods. NOMAD-EEG is a 12.0 M-parameter diffusion model that alternates state-space layers along time, without patches, with attention across electrodes identified only by their 3D scalp position. It was pre-trained on 147 h of unlabelled EEG from a single corpus and evaluated with frozen linear probes on ten datasets with 1 to 64 electrodes against seven foundation models run under one protocol, and on two unseen populations. Results. Despite its small size (12.0 M parameters against 69 M for REVE) and single pre- training corpus, NOMAD-EEG outperformed all seven foundation models on EEGMMIDB mo- tor imagery (balanced accuracy 0.442), Siena seizure detection (macro F1 0.622) and Bonn seizure detection (AUROC 0.999), had the highest accuracy of the seven compared models on TUEP epilepsy diagnosis (0.640) and, after fine-tuning, the highest \k{appa} (0.635) and weighted F1 (0.814) on TUEV event classification. It was the best of five models on P300 detection in autism with an unseen eight-electrode montage (AUROC 0.718). With only 5 of 19 electrodes, its mean AUROC fell only from 0.942 to 0.891. Conclusion. Identifying electrodes by position and modelling time without patches let one small model, pre-trained on one corpus, outperform much larger EEG foundation models on seizure detection, epileptiform event classification, epilepsy diagnosis and motor imagery.
Figures & tables
Model
Objective
Channel identity
Any set?
Order-free?
Time handling
NOMAD-EEG (ours)
Diffusion
3D coordinates (Fourier), every block
Yes
Yes
20 ms steps, no patches, no length cap
REVE [ 12 ]
Masked AE
3D coordinates + patch index
Yes
Yes
1 s patches, full attention
LUNA [ 13 ]
Masked recon.
3D coordinates (NeRF-style)
Yes
Yes
Patches; channels compressed to queries
BrainOmni [ 14 ]
Tokeniser + masked
Sensor position, orientation, type
Yes
Yes
Windows; channels compressed
DIVER-0 [ 25 ]
Masked recon.
Attention bias, no coordinates
Yes
Yes
1 s patches
LaBraM [ 2 ]
Masked codes
Learned embedding per name
Vocabulary
Yes
1 s patches, ≤ 16 patches
Table 1: How EEG foundation models identify channels and handle time. Entries for the eight models we run were checked against their released code; the rest are from the papers. “Any electrode set” means the model accepts channels that were not in its pre-training montage without retraining.
Figure 1: One model, any montage. The same pretrained NOMAD-EEG weights read five electrode layouts used in this work, from a single electrode to a 64-electrode cap. Electrode positions are the 3D coordinates the network receives, projected to 2D (nose up). Blue: positions present in the pre-training montage. Red: positions never seen in pre-training, or the midpoint used for a bipolar derivation (open circles mark its two electrodes).
Figure 2: Overview of NOMAD-EEG. (a) Pre-training. A clean 30 s TUEV window x0 (21 electrodes, 200 Hz) is corrupted with Gaussian noise ϵ following the cosine schedule, and the network, given the electrode positions pe , the noise level t and the corpus c , predicts the noise. It consists of a convolutional stem shared by all electrodes (one 256-dimensional feature every 20 ms), 16 identical blocks, and an output head that maps the summed skip outputs back to the sample rate. The flow-matching variant differs only in the objective. (b) One block. The signal is a grid of tokens with one row per electrode and one column per 20 ms step. After layer normalisation and FiLM conditioning on (t,c) , a bidirectional S4D layer filters each row along time with a kernel that spans the whole segment and is shared by all electrodes (blue); attention then mixes each column across electrodes, every token carrying Fourier features ϕ(pe) of its electrode’s 3D position (red). Because electrodes are identified only by position, the same block reads any electrode set, for example the 21-electrode pre-training montage or two bipolar channels placed at the midpoints of their electrode pairs. A gated activation gives the output u , which feeds the next block and the skip sum. (c) Downstream, the input is clean and the gates of blocks 4, 8, 12 and 16 are read. The frozen probe averages them over electrodes and keeps the mean and standard deviation in five time windows (10,240 features) for a linear classifier; the fine-tuning head pools E×25 tokens by attention from a learned query ( Fig. 4 ). Waveforms, attention weights and highlighted tokens are illustrative; electrode positions are the coordinates the network receives.
Figure 3: Factorised mixing. Inside the network an EEG segment is a grid with one row per electrode and one column per 20 ms step. Each block mixes information along every row with a bidirectional S4D layer and then down every column with attention across electrodes. Neither operation has a parameter whose size depends on the number of rows or columns, so the grid can have any shape.
Figure 4: Downstream task heads. (a) The frozen feature extractor averages the gate activations over electrodes, splits time into five equal windows and keeps the mean and standard deviation in each, giving 10,240 features whatever the montage or length. (b) The fine-tuning head keeps electrodes separate: each electrode in each of 25 windows is a token, and a learned query pools the tokens by attention. Traces and attention shading are illustrative.
Dataset
Task
Classes
Channels
Segment
Train / val / test
TUAB
abnormal EEG
2
21
5 s
100,000 / 20,000 / 50,000
TUSZ
seizure
2
21
5 s
100,000 / 20,000 / 50,000
TUEV
EEG events
6
21
5 s
83,932 (5-fold CV) / 29,421
TUEP
epilepsy diagnosis
2
19
5 s
12,000 from 200 patients (5-fold CV)
Siena
seizure
2
21
5 s
74,512 / 8,279 / 18,729
Bonn
seizure
2
1
5 s
360 / 40 / 100
Table 2: Downstream datasets. Train/val/test counts are segments; all splits are subject-disjoint except TUAB, TUSZ, Siena and Bonn, which reuse the fixed splits of our earlier benchmark, and SEED-V, which follows the CBraMod protocol (film clips 1–5, 6–10 and 11–15 of every session; all 16 subjects in every split). TUEP is evaluated by 5-fold cross-validation over its 200 patients.
NOMAD-EEG (Base)
NOMAD-EEG-Large
Parameters
12.0 M
67.6 M
Width / blocks / heads
256 / 16 / 4
512 / 24 / 8
Feature dimension
10,240
20,480
Pre-training data
147 h, 1 corpus
6,984 h, 7 corpora
Montages
1 (21 electrodes)
38 channel groups, 1–62 electrodes
Electrodes per sample
21
random subset of 2–32
Table 3: Base and Large configurations. An epoch is as many samples as the training set has 30 s windows; for Large, whose samples are crops of at most 30 s, it is an upper bound on the signal seen.
Figure 5: Summary of the epilepsy tasks. Macro F1 of the frozen linear probe of every model on seizure detection (TUSZ, Siena, Bonn), epileptiform event classification (TUEV, mean of the six per-class F1 values) and epilepsy diagnosis (TUEP, patient level). Macro F1 is shown because it measures decisions at the classifier’s threshold, which in seizure detection determines the false alarms; the primary metric of each dataset is given in its table. Red dots: the model’s pre-training data contained that dataset ( Section 2.8 ); LaBraM’s TUEP value is shown for completeness although it is excluded from Table 7 for that reason. Values come from the same probes as Tables 5 , 8 , 9 , 6 and 7 ; the TUEV macro F1 is computed from the per-class F1 values of those predictions.
Model
Dim
AUROC
PR-AUC
Bal. acc.
Macro F1
NOMAD-EEG (ours)
10,240
0.838 [0.834, 0.842]
0.820
0.770
0.771
fine-tuned
—
0.833 [0.830, 0.837]
0.785
0.770
0.771
EEGDM [ 9 ]
2,560
0.869 [0.865, 0.872]
0.862
0.792
0.792
REVE [ 12 ]
512
0.870 [0.867, 0.873]
0.880
0.800
0.802
CBraMod [ 5 ]
200
0.808 [0.804, 0.812]
0.803
0.738
0.739
LaBraM [ 2 ]
200
0.788 [0.784, 0.792]
0.785
0.713
0.714
Table 4: TUAB abnormal-EEG detection (5 s windows; 50,000 test windows), frozen linear probe. Brackets: 95% bootstrap CI. TUAB’s training split was in BIOT’s pre-training data, and CBraMod and EEGMamba were pre-trained on the TUH corpus it is drawn from ( Section 2.8 ). The indented row is NOMAD-EEG fine-tuned end to end (not part of the bold comparison); its epoch was selected on the benchmark’s validation windows, which come from the training patients ( Section 2.9 ).
Model
Dim
AUROC
PR-AUC
Bal. acc.
Macro F1
NOMAD-EEG (ours)
10,240
0.866 [0.860, 0.873]
0.446
0.787
0.620
fine-tuned
—
0.854 [0.847, 0.860]
0.345
0.773
0.676
EEGDM [ 9 ]
2,560
0.874 [0.866, 0.881]
0.527
0.801
0.642
REVE [ 12 ]
512
0.879 [0.873, 0.886]
0.463
0.798
0.611
CBraMod [ 5 ]
200
0.855 [0.847, 0.862]
0.455
0.776
0.586
LaBraM [ 2 ]
200
0.845 [0.837, 0.851]
0.416
0.765
0.589
Table 5: TUSZ seizure detection (5 s windows; 50,000 test windows at the true seizure prevalence), frozen linear probe. TUSZ was in LaBraM’s and DiffEEG’s pre-training data, and REVE, CBraMod and EEGMamba were pre-trained on the TUH corpus it is drawn from ( Section 2.8 ). The indented row is NOMAD-EEG fine-tuned end to end (not part of the bold comparison); its epoch was selected on the benchmark’s validation windows, which come from the training patients ( Section 2.9 ).
Model
Dim
Bal. acc.
Cohen κ
Weighted F1
NOMAD-EEG (ours)
10,240
0.542 [0.529, 0.554]
0.465
0.719
fine-tuned
—
0.597 [0.588, 0.605]
0.635
0.814
EEGDM [ 9 ]
2,560
0.647 [0.638, 0.657]
0.568
0.783
REVE [ 12 ]
512
0.565 [0.553, 0.577]
0.515
0.753
CBraMod [ 5 ]
200
0.466 [0.455, 0.477]
0.346
0.635
LaBraM [ 2 ]
200
0.521 [0.510, 0.534]
0.365
0.633
Table 6: TUEV event classification (6 classes; 29,421 test events; in the pre-training data of NOMAD-EEG, EEGDM and BIOT, and in the TUH corpus on which CBraMod and EEGMamba were pre-trained, Section 2.8 ), frozen linear probe with 5-fold subject-grouped cross-validation. The indented row is NOMAD-EEG fine-tuned end to end; bold marks the best value in each column including this row, although the baselines were evaluated only as frozen probes.
Model
Dim
AUROC
PR-AUC
Accuracy
Weighted F1
Macro F1
NOMAD-EEG (ours)
10,240
0.679 [0.607, 0.751]
0.675
0.640
0.638
0.638
EEGDM [ 9 ]
2,560
0.673 [0.599, 0.745]
0.638
0.610
0.609
0.609
REVE [ 12 ] ‡
512
0.671 [0.596, 0.748]
0.674
0.605
0.605
0.605
CBraMod [ 5 ] ‡
200
0.671 [0.591, 0.745]
0.667
0.615
0.615
0.615
BIOT [ 4 ]
256
0.675 [0.605, 0.748]
0.633
0.630
0.630
0.630
EEGMamba [ 6 ] ‡
200
0.694 [0.619, 0.766]
0.701
0.625
0.625
0.625
Table 7: TUEP epilepsy diagnosis (200 patients, 100 with and 100 without epilepsy; 19 electrodes, 60 random 5 s windows per patient), frozen encoders and a linear probe evaluated on unseen patients (5-fold cross-validation stratified by diagnosis, identical folds for every model). A patient’s score is the mean probability of their windows. Brackets: 95% CI from 1,000 bootstrap resamples of patients; accuracy and F1 at a 0.5 threshold. LaBraM is not included because TUEP was part of its pre-training data (under the same protocol its patient-level AUROC is 0.741 and accuracy 0.680). ‡ The model was pre-trained on the TUH corpus from which TUEP is drawn ( Section 2.8 ). Bold: best among the models whose pre-training did not include TUEP recordings (no mark); underlined: best overall.
Model
Dim
AUROC
PR-AUC
Accuracy
Weighted F1
Macro F1
NOMAD-EEG (ours)
10,240
0.882 [0.847, 0.914]
0.306
0.979
0.984
0.622
EEGDM [ 9 ]
2,560
0.915 [0.889, 0.938]
0.248
0.916
0.950
0.534
REVE [ 12 ]
512
0.906 [0.872, 0.937]
0.460
0.946
0.966
0.562
CBraMod [ 5 ]
200
0.850 [0.806, 0.888]
0.189
0.872
0.925
0.502
LaBraM [ 2 ]
200
0.797 [0.756, 0.836]
0.056
0.833
0.903
0.478
BIOT [ 4 ]
256
0.707 [0.657, 0.755]
0.020
0.780
0.870
0.455
Table 8: Siena seizure detection (external to NOMAD-EEG; 18,729 test windows, 0.6% seizure in training), frozen linear probe. Siena recordings were in REVE’s, LaBraM’s and EEGMamba’s pre-training data ( Section 2.8 ).
Model
Dim
AUROC
PR-AUC
Bal. acc.
Macro F1
NOMAD-EEG (ours)
10,240
0.999 [0.996, 1.000]
0.998
0.994
0.985
single electrode only
10,240
0.995 [0.981, 1.000]
0.983
0.944
0.913
EEGDM [ 9 ]
2,560
0.994 [0.982, 1.000]
0.975
0.963
0.915
REVE [ 12 ]
512
0.998 [0.991, 1.000]
0.992
0.994
0.985
CBraMod [ 5 ]
200
0.996 [0.986, 1.000]
0.985
0.956
0.903
LaBraM [ 2 ]
200
0.995 [0.983, 1.000]
0.982
0.950
0.890
Table 9: Bonn seizure detection (external; single-channel recordings; 100 test segments), frozen linear probe. All models read the recording in a 22-channel frame; the indented row feeds NOMAD-EEG the one signal-carrying electrode alone.
Model
Dim
AUROC
PR-AUC
Bal. acc.
Macro F1
NOMAD-EEG (ours)
10,240
0.942 [0.929, 0.954]
0.951
0.832
0.831
fine-tuned
—
0.872 [0.851, 0.893]
0.792
0.821
0.821
EEGDM [ 9 ]
2,560
0.946 [0.935, 0.958]
0.945
0.857
0.856
REVE [ 12 ]
512
0.960 [0.951, 0.968]
0.962
0.842
0.841
CBraMod [ 5 ]
200
0.860 [0.838, 0.883]
0.810
0.804
0.803
LaBraM [ 2 ]
200
0.904 [0.887, 0.922]
0.886
0.838
0.837
Table 10: Mumtaz depression vs. healthy (external; 19 electrodes, 5 s windows; 1,179 test windows from 10 held-out subjects), frozen linear probe. The indented row is NOMAD-EEG fine-tuned end to end (no baseline was fine-tuned).
Model
Dim
Bal. acc.
Cohen κ
Weighted F1
NOMAD-EEG (ours)
10,240
0.442 [0.419, 0.463]
0.256
0.442
fine-tuned
—
0.596 [0.574, 0.618]
0.461
0.599
EEGDM [ 9 ]
2,560
0.347 [0.325, 0.367]
0.130
0.338
REVE [ 12 ]
512
0.371 [0.348, 0.393]
0.161
0.370
CBraMod [ 5 ]
200
0.308 [0.287, 0.330]
0.078
0.306
LaBraM [ 2 ]
200
0.302 [0.280, 0.324]
0.070
0.301
Table 11: EEGMMIDB motor imagery (external to NOMAD-EEG; 4 classes, 64 electrodes, 4 s trials; 1,803 test trials from 20 held-out subjects), frozen linear probe. EEGMMIDB was in LaBraM’s pre-training data ( Section 2.8 ). The indented row is NOMAD-EEG fine-tuned end to end.
Model
Dim
Bal. acc.
Cohen κ
Weighted F1
NOMAD-EEG (ours)
10,240
0.757 [0.748, 0.766]
0.680
0.770
fine-tuned
—
0.758 [0.748, 0.768]
0.694
0.784
EEGDM [ 9 ]
2,560
0.779 [0.770, 0.788]
0.718
0.802
REVE [ 12 ]
512
0.760 [0.750, 0.770]
0.675
0.771
CBraMod [ 5 ]
200
0.725 [0.715, 0.735]
0.639
0.743
LaBraM [ 2 ]
200
0.738 [0.729, 0.747]
0.676
0.770
Table 12: Sleep-EDF-20 sleep staging (external; 5 stages, 2 bipolar channels, 30 s epochs; 9,137 test epochs from 4 held-out subjects), frozen linear probe. The indented row is NOMAD-EEG fine-tuned end to end.
NOMAD-EEG
Bal. acc.
Cohen κ
Weighted F1
frozen linear probe
0.315 [0.310, 0.319]
0.136
0.302
fine-tuned
0.384 [0.380, 0.389]
0.231
0.392
Table 13: SEED-V emotion recognition (external; 5 emotions, 59 electrodes, 1 s windows; within every session film clips 1–5 train, 6–10 validation, 11–15 test, all 16 subjects in every split), NOMAD-EEG only. Chance balanced accuracy is 0.20.
Model
Channels read
AUROC
Bal. acc.
PR-AUC
Object acc. (1/3/10 runs)
Per-subject AUROC range
NOMAD-EEG
8
0.718 [0.695, 0.739]
0.660
0.289
32.6 / 51.5 / 75.5%
0.652–0.805
EEGDM
4 bipolar
0.667 [0.631, 0.699]
0.624
0.227
27.8 / 42.4 / 61.2%
0.540–0.805
REVE
8
0.651 [0.626, 0.675]
0.612
0.209
24.6 / 37.9 / 57.5%
0.563–0.752
BIOT
2 bipolar
0.530 [0.521, 0.539]
0.520
0.136
14.5 / 17.2 / 21.8%
0.503–0.571
DiffEEG
6
0.522 [0.508, 0.534]
0.517
0.133
13.9 / 15.1 / 19.4%
0.457–0.556
Table 14: P300 detection on BCIAUT-P300 (15 participants with ASD, 168,000 epochs, 12.5% targets). Encoders are frozen and never trained on this corpus; a linear probe is evaluated on subjects it never saw (five folds of three subjects). Brackets: 95% CI from a bootstrap over subjects. “Object acc.” is the rate at which the attended object among eight is selected from the first 1, 3 or 10 runs of a block (chance 12.5%).
Model
Window AUROC
Participant AUROC
Participant bal. acc.
NOMAD-EEG (ours)
0.679 [0.609, 0.754]
0.684 [0.590, 0.777]
0.620 [0.542, 0.706]
EEGDM [ 9 ]
0.783 [0.709, 0.845]
0.795 [0.714, 0.868]
0.702 [0.618, 0.782]
REVE [ 12 ]
0.653 [0.571, 0.733]
0.643 [0.547, 0.740]
0.626 [0.543, 0.705]
BIOT [ 4 ]
0.659 [0.583, 0.730]
0.681 [0.593, 0.771]
0.672 [0.593, 0.748]
DiffEEG [ 8 ]
0.634 [0.585, 0.683]
0.713 [0.623, 0.794]
0.672 [0.587, 0.751]
Table 15: Parkinson’s disease vs. control (OpenNeuro ds004584; 149 participants, 100 PD / 49 controls; 60 electrodes, 5 s windows), frozen encoders and a linear probe evaluated on unseen participants (5-fold cross-validation stratified by group, identical folds for every model). A participant’s score is the mean probability of their windows. Brackets: 95% CI from 1,000 bootstrap resamples of participants.
Dataset
Setting
Metric
Pre-trained
Random init.
Δ
TUEV
linear probe
Bal. acc.
0.542 [0.529, 0.554]
0.434 [0.422, 0.446]
+0.107
TUEV
fine-tuned
Bal. acc.
0.597 [0.588, 0.605]
0.465 [0.455, 0.474]
+0.132
TUAB
linear probe
AUROC
0.838 [0.834, 0.842]
0.798 [0.794, 0.801]
+0.040
Siena
linear probe
AUROC
0.882 [0.847, 0.914]
0.892 [0.858, 0.924]
-0.010
Siena
linear probe
Macro F1
0.622 [0.596, 0.646]
0.551 [0.538, 0.563]
+0.071
Siena
linear probe
PR-AUC
0.306 [0.221, 0.392]
0.192 [0.134, 0.262]
+0.114
Table 16: Effect of pre-training: NOMAD-EEG with its pre-trained weights versus the same architecture with random weights, under the identical feature extraction and protocol. Brackets: 95% bootstrap CI. Δ in bold where the two CIs do not overlap.
Dataset
Metric
Diffusion
Flow
Flow − diffusion
TUEV
Bal. acc.
0.542
0.486
-0.055 [-0.067, -0.044]
TUAB
AUROC
0.838
0.827
-0.011 [-0.013, -0.008]
TUAB
Bal. acc.
0.770
0.762
-0.008 [-0.011, -0.004]
TUSZ
AUROC
0.866
0.870
+0.004 [-0.002, 0.010]
TUSZ
Bal. acc.
0.787
0.798
+0.011 [0.003, 0.018]
Siena
AUROC
0.882
0.866
-0.016 [-0.037, 0.004]
Table 17: Diffusion versus flow matching. Both models share architecture, data, number of steps, optimiser and feature extraction; only the objective differs. Frozen linear probe on each dataset; “Flow − diffusion” is the paired difference on the same test samples with a 95% CI from 2,000 paired bootstrap resamples.
Figure 6: Missing-electrode test on Mumtaz (1,179 test windows). The linear probe is trained once on the full montage and applied unchanged. (a) Each scalp shows one test condition: filled dots are the electrodes given to the model, hollow dots those removed; the shading and the label give the AUROC, and BAcc is the balanced accuracy at the probe’s threshold. Rows: 19, 14, 10 and 5 electrodes; columns: five random subsets per size. (b) The same results against the number of electrodes given: dots are individual subsets, lines the mean, bands the range over subsets. AUROC stays high as electrodes are removed, while balanced accuracy drops because the scores shift relative to the threshold fitted on the full montage.
Dataset
Metric
In Large
Base
Large
Large
pre-training
step 27,000
step 51,000
Bonn
AUROC
no
0.999 [0.996, 1.000]
0.999 [0.994, 1.000]
1.000 [1.000, 1.000]
TUSZ
AUROC
training patients
0.866 [0.860, 0.873]
0.878 [0.872, 0.884]
0.856 [0.849, 0.862]
TUAB
AUROC
training patients
0.838 [0.834, 0.842]
0.828 ↓ [0.824, 0.831]
0.837 [0.834, 0.841]
Siena
AUROC
no
0.882 [0.847, 0.914]
0.867 [0.822, 0.906]
0.867 [0.823, 0.907]
Mumtaz
AUROC
no
0.942 [0.929, 0.954]
0.913 ↓ [0.895, 0.929]
0.913 [0.894, 0.929]
Table 18: Scaling model size and pre-training data together: NOMAD-EEG (Base, 12.0 M parameters, 147 h of TUEV) against NOMAD-EEG-Large (67.6 M parameters, 6,984 h from seven corpora) at two points of its pre-training, under the identical frozen linear probe and splits used for every other table. Each dataset’s primary metric with its 95% bootstrap CI; bold marks the best value in each row. ↓ / ↑ : the Large CI lies entirely below/above Base’s (non-overlapping intervals, not a paired test). In Large pre-training : whether that dataset’s recordings were in Large’s unlabelled pre-training data. p patient level. b TUEV is also Base’s whole pre-training corpus.
Figure 7: Where NOMAD-EEG finds evidence of a seizure. Two 5 s segments from the TUSZ test set, scored by the frozen NOMAD-EEG with the TUSZ linear probe of Table 5 . The segments were chosen by a rule fixed in advance: a random (seeded) test window labelled seizure that the model classifies with probability above 0.9, and a random normal window classified with probability below 0.1, i.e. typical correct decisions rather than hand-picked ones. Shading behind each trace: positive Integrated-Gradients attribution (evidence for seizure) per electrode and 0.25 s; bottom: its sum over electrodes. Both panels share the colour and amplitude scales. The figure illustrates single decisions and is not a quantitative test of seizure localisation.
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31× larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.
Kuntal Kokate, Bruno Aristimunha, Dung Truong +1
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France
Foundation models (FMs) promise to extract unified representations that generalize across downstream tasks. They have emerged across fields, including electroencephalography (EEG), but it is less clear how effective they are in this particular field. Published evaluations differ in datasets, in the EEG-specific preprocessing that might influence reported results, and in the reported metrics, frequently obscuring the clinical relevance in EEG. We introduce NeuroAtlas, the largest EEG benchmark to date: 42 datasets and 260k hours covering clinical EEG (epilepsy, sleep medicine, brain age estimation) and brain-computer interfaces, and include multiple datasets per task along with bespoke clinical evaluation metrics. Besides evaluating EEG-FMs with respect to supervised baselines, we present results from generic time-series FMs. We report three findings. First, EEG-specific FMs do not consistently outperform time-series FMs, which have neither EEG-focused architectures nor been pretrained on EEG. Second, standard machine learning metrics are insufficient to assess clinical utility: thus, we thoroughly evaluate more appropriate measures such as the quality of event-level decision-making, hypnogram-derived features, and the brain-age gap in the domains of epilepsy, sleep, and brain age, respectively. Third, model rankings and performance can vary substantially within domains. We conclude that pretrained models perform largely on par, with only narrow advantages for a few, and that current models do not yet deliver on the promise of an out-of-the-box unified EEG model. NeuroAtlas exposes this gap and provides the datasets and metrics for the next generation of unified EEG FMs.
Konstantinos Kontras, Trui Osselaer, Stylianos G. Mouslech +12
Electroencephalography (EEG) is a critical tool for monitoring brain activity and diagnosing neurological disorders such as epilepsy. However, learning meaningful representations from raw EEG signals remains challenging due to limited annotations, substantial inter-subject variability, and complex temporal dynamics. Recent EEG foundation models (FMs) have demonstrated promising performance through transformer-based architectures and large-scale self-supervised pretraining, yet they often incur substantial computational costs and exhibit diminishing returns with increasing model and dataset scale, limiting their practicality in clinical settings. To address these challenges, we propose EEGDM, a diffusion-based EEG representation learning framework. Specifically, EEGDM introduces a Structured State-Space Model for Diffusion Pretraining (SSMDP) that effectively captures long-range temporal dependencies through generative diffusion training on unlabeled EEG data. The representations are subsequently leveraged for downstream tasks via our Latent Fusion Module (LFM), which integrates multi-layer latent features of SSMDP. We evaluate EEGDM on three EEG benchmarks spanning EEG event classification (TUEV), seizure detection (CHB-MIT), and seizure classification (IIIC). Compared with existing state-of-the-art methods, including EEG FMs, EEGDM achieves competitive performance across diverse tasks and datasets, exceeding existing methods in most cases while requiring substantially fewer samples for both pretraining and downstream adaptation. These results demonstrate that diffusion-based learning can unlock more discriminative EEG representations, with direct implications for epilepsy diagnosis and management. Our source code and pretrained checkpoints are publicly available at: https://github.com/jhpuah/EEGDM.
Jia Hong Puah, Sim Kuan Goh, Ziwei Zhang +6
School of Artificial Intelligence and Robotics, Xiamen University Malaysia · School of Energy and Chemical Engineering, Xiamen University Malaysia · Department of Biomedical Engineering, Universiti Malaya +3