Core-matrix protocol · Transfer to a new person

Sleep staging on EESM19 scalp subset

The released protocol, with the score of every method run under it.

The protocol at a glance

Task
Five stages · balanced scalp-EEG sample
Evaluation
Transfer to a new person
People
20
Data
3,000 epochs · 5 participant-disjoint folds
Input
6 channels · 30-second windows
Chance level
20.0%
Metrics
Balanced accuracy (primary) · Macro F1 (secondary)

Results

Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level.

Spectral ridge
72.2% (70.7%–73.8%)
LaBraM
71.1% (69.0%–73.2%)
CBraMod
71.3% (68.7%–73.7%)
EEGNet
56.5% (54.9%–58.3%)
ShallowFBCSPNet
55.0% (53.1%–56.8%)
Deep4Net
45.2% (42.7%–47.7%)
CSP+LDA
50.1% (46.9%–53.1%)
Balanced accuracy, every method under this protocol. Dot: the estimate; line: descriptive 95% interval; dashed line: chance level.
MethodTraining modeBalanced accuracyMacro F1ChannelsPeopleScoring time
Spectral ridgeClassical methodSupervised fit72.2%70.7%–73.8%0.705Mean across held-out participants6202.6 s
LaBraMFoundation modelFrozen encoder + ridge head71.1%69.0%–73.2%0.697Mean across held-out participants62011.5 s
CBraModFoundation modelFrozen encoder + ridge head71.3%68.7%–73.7%0.702Mean across held-out participants62010.2 s
EEGNetCompact modelScratch · 10 epochs56.5%54.9%–58.3%0.545Mean across held-out participants62090.4 s
ShallowFBCSPNetCompact model · matrix recipe · AdamW · Local Ubuntu / CUDAScratch · 10 epochs55.0%53.1%–56.8%0.519Mean across held-out participants620—
Deep4NetCompact model · matrix recipe · AdamW · Local Ubuntu / CUDAScratch · 10 epochs45.2%42.7%–47.7%0.419Mean across held-out participants620—
CSP+LDAClassical method · one-vs-rest CSP · shrinkage LDA · Local Ubuntu / CPUSupervised fit50.1%46.9%–53.1%0.483Mean across held-out participants620—

Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106.

Not applicable under this protocol: Standard CCA, FBCCA (they need the stimulation frequencies, and this protocol has none); Temporal ridge, XDAWNCov+TS+LR, xDAWN+LDA (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one).

Not run under this protocol: TS+LR, MDM.

Pick-two protocol comparison

Compare any two methods on Sleep staging

Inspect point estimates and 95% intervals side by side on a shared axis. Paired differences appear only where a paired interval is published. Dropdowns are grouped by family alphabetically — never ordered by score.

0%25%50%75%100%Chance 20.0%ShallowFBCSPNet55.0% [53.1–56.8]EEGNet56.5% [54.9–58.3]
Published paired contrastShallowFBCSPNet − EEGNet: −1.6 pp[−3.4, +0.4 pp]

95% interval, unadjusted · n = 20 participants · Paired interval includes zero

against the matrix's one-seed EEGNet cell; training variance not included

Default view (Sleep staging): ShallowFBCSPNet balanced accuracy 55.0% (95% interval 53.1–56.8%); EEGNet 56.5% (95% interval 54.9–58.3%); chance 20.0%.

Published paired contrasts table (8 pairs)
Pair (left minus right)Mean differenceGoverning intervalParticipantsAdjustment & notes
CSP+LDA − EEGNet−6.4 pp[−9.3, −3.3 pp]2095% interval, unadjustedagainst the matrix's one-seed EEGNet cell; training variance not included
Deep4Net − EEGNet−11.3 pp[−13.6, −9.0 pp]2095% interval, unadjustedagainst the matrix's one-seed EEGNet cell; training variance not included
ShallowFBCSPNet − EEGNet−1.6 pp[−3.4, +0.4 pp]2095% interval, unadjustedagainst the matrix's one-seed EEGNet cell; training variance not included
ShallowFBCSPNet − Deep4Net+9.7 pp[+7.0, +12.7 pp]2095% interval
Log band-power ridge − Spectral ridge−6.2 pp[−8.5, −3.9 pp]2095% paired participant bootstrap interval, unadjusted
LaBraM at Scale Quark-L − LaBraM at Scale Quark-B−1.8 pp[−3.3, −0.3 pp]2095% paired participant bootstrap interval, unadjusted (21 ladder contrasts)
LaBraM at Scale Quark-H − LaBraM at Scale Quark-L+2.5 pp[+1.0, +4.2 pp]2095% paired participant bootstrap interval, unadjusted (21 ladder contrasts)
LaBraM at Scale Quark-H − LaBraM at Scale Quark-B+0.7 pp[−0.9, +2.5 pp]2095% paired participant bootstrap interval, unadjusted (21 ladder contrasts)

How these rows were run

ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe

Recipe
ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW; CSP+LDA: one-vs-rest CSP · shrinkage LDA
Hardware
ShallowFBCSPNet, Deep4Net: Local Ubuntu / CUDA; CSP+LDA: Local Ubuntu / CPU

Paired differences

Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people.

ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe

against the matrix's one-seed EEGNet cell; training variance not included
DifferenceMean (pp)95% interval, unadjustedBonferroni interval over 18 contrastsReading
CSP+LDA − EEGNetCSP+LDA (one-vs-rest CSP)−6.4 pp−9.3 to −3.3−10.8 to −1.5lower
Deep4Net − EEGNet−11.3 pp−13.6 to −9.0−14.7 to −7.8lower
ShallowFBCSPNet − EEGNet−1.6 pp−3.4 to 0.4−4.7 to 1.5unresolved

The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance.

Re-training on the CUDA card is not bit-exact; re-training the published semantic-target EEGNet cell there moved it by 0.1 pp.

Beside the core matrix

Log band-power ridge (5 bands, visible channels)

Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above.

MethodTraining modeBalanced accuracyMacro F1ChannelsPeople
Spectral ridgeClassical method · core matrix row aboveSupervised fit72.23%70.70%–73.80%0.705620
Log band-power ridge (5 bands, visible channels)Classical methodSupervised fit66.00%63.03%–68.93%0.636620

Differences from two published rows

Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction.

Minus the Spectral ridge row: −6.23 pp (−8.50 to −3.90): below. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe.

Minus ST-EEGFormer Large, the highest of the 14 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −13.60 pp (−15.80 to −11.37): below. Choosing it on the test folds leans the comparison against this row.

Beside the core matrix

Further foundation encoders, frozen

Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context.

Spectral ridge
72.2% (70.7%–73.8%)
LaBraM
71.1% (69.0%–73.2%)
CBraMod
71.3% (68.7%–73.7%)
EEGNet
56.5% (54.9%–58.3%)
REVE Base
74.6% (72.9%–76.4%)
REVE Large
76.6% (75.2%–78.0%)
LUNA Base
67.5% (65.3%–69.5%)
LUNA Large
72.4% (70.4%–74.0%)
BrainOmni Base
67.7% (65.8%–69.5%)
CodeBrain
70.7% (68.8%–72.7%)
EEGMamba
64.1% (61.6%–66.4%)
ST-EEGFormer Base
77.3% (75.6%–78.8%)
ST-EEGFormer Large
79.6% (78.2%–81.0%)
eeg-fm-masking MAE (r=9 cm, L=2)
67.9% (65.4%–70.4%)
ERP-FM Base
71.0% (69.5%–72.6%)
SingLEM
35.3% (33.7%–37.0%)
ZUNA 1.1
60.1% (57.8%–62.3%)
Balanced accuracy under this protocol. Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level.
EncoderTraining modeBalanced accuracyMacro F1ChannelsPeoplePretraining exposure
REVE Base1Foundation modelFrozen encoder + ridge head74.6%72.9%–76.4%0.733620not in the authors' published pretraining list (checked 2026-10-04)source ↗
REVE Large2Foundation modelFrozen encoder + ridge head76.6%75.2%–78.0%0.760620not in the authors' published pretraining list (checked 2026-10-04)source ↗
LUNA Base3Foundation modelFrozen encoder + ridge head67.5%65.3%–69.5%0.656620not in the authors' published pretraining list (checked 2026-10-04)source ↗
LUNA Large4Foundation modelFrozen encoder + ridge head72.4%70.4%–74.0%0.707620not in the authors' published pretraining list (checked 2026-10-04)source ↗
BrainOmni Base5Foundation modelFrozen encoder + ridge head67.7%65.8%–69.5%0.674620not in the authors' published pretraining list (checked 2026-10-04)source ↗
CodeBrain6Foundation modelFrozen encoder + ridge head70.7%68.8%–72.7%0.694620not in the authors' published pretraining list (checked 2026-10-04)source ↗
EEGMamba7Foundation modelFrozen encoder + ridge head64.1%61.6%–66.4%0.625620not in the authors' published pretraining list (checked 2026-10-04)source ↗
ST-EEGFormer Base8Foundation modelFrozen encoder + ridge head77.3%75.6%–78.8%0.765620not in the authors' published pretraining list (checked 2026-10-04)source ↗
ST-EEGFormer Large9Foundation modelFrozen encoder + ridge head79.6%78.2%–81.0%0.789620not in the authors' published pretraining list (checked 2026-10-04)source ↗
eeg-fm-masking MAE (r=9 cm, L=2)10Foundation modelFrozen encoder + ridge head67.9%65.4%–70.4%0.668620not in the authors' published pretraining list (checked 2026-10-04)source ↗
ERP-FM Base11Foundation modelFrozen encoder + ridge head71.0%69.5%–72.6%0.695620not in the authors' published pretraining list (checked 2026-10-04)source ↗
SingLEM12Foundation modelFrozen encoder + ridge head35.3%33.7%–37.0%0.325620not in the authors' published pretraining list (checked 2026-10-04)source ↗
ZUNA 1.113Foundation modelModel card: research use only, not for diagnosis or clinical use.Frozen encoder + ridge head60.1%57.8%–62.3%0.579620Exposure unknownunknown: the authors do not list their pretraining data closely enough to decidesource ↗

Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence.

Row footnotes

  1. Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower.
  2. Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower.
  3. Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime.
  4. As LUNA Base. Frozen probes only.
  5. p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages.
  6. Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream).
  7. Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets).
  8. BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus).
  9. BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only.
  10. Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel.
  11. In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline.
  12. semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer.
  13. Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction.
  14. A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry.
Masking ablation: three sibling checkpoints of eeg-fm-masking

Intervals are marginal: the same people under two checkpoints are not tested as a pair.

EncoderTraining modeBalanced accuracyMacro F1ChannelsPeoplePretraining exposure
eeg-fm-masking MAE (r=9 cm, L=2)10Foundation model · matrix row aboveFrozen encoder + ridge head67.9%65.4%–70.4%0.668620not in the authors' published pretraining list (checked 2026-10-04)source ↗
eeg-fm-masking JEPA (r=9 cm, L=2)14Foundation modelFrozen encoder + ridge head70.3%68.0%–72.7%0.695620not in the authors' published pretraining list (checked 2026-10-04)source ↗
eeg-fm-masking MAE (r=one channel, L=1)14Foundation modelFrozen encoder + ridge head65.2%63.0%–67.3%0.644620not in the authors' published pretraining list (checked 2026-10-04)source ↗
eeg-fm-masking JEPA (r=one channel, L=1)14Foundation modelFrozen encoder + ridge head68.3%65.7%–70.8%0.675620not in the authors' published pretraining list (checked 2026-10-04)source ↗

Weights licences and terms

No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared.

Notes on these rows

Beside the core matrix

EEGPT, frozen

EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring.

Spectral ridge
72.2% (70.7%–73.8%)
LaBraM
71.1% (69.0%–73.2%)
CBraMod
71.3% (68.7%–73.7%)
EEGNet
56.5% (54.9%–58.3%)
EEGPT
75.5% (73.9%–77.1%)
Balanced accuracy under this protocol. Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level.
EncoderTraining modeBalanced accuracyMacro F1ChannelsPeoplePretraining exposure
EEGPTFoundation modelFrozen encoder + ridge head75.5%73.9%–77.1%0.748620not in the authors' published pretraining list (checked 2026-10-12)source ↗

Footnotes to this row

  1. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average.
  2. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with.

Against the published rows

Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction):

EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows.

Licence and terms

Frozen-probe comparison

LaBraM at Scale and REVE Small, frozen

Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control.

Balanced accuracy with its 95% interval, and macro F1, frozen encoders.
EncoderBalanced accuracy (95% interval)Macro F1Exposure
LaBraM at Scale · Quark ladder
Quark-BLocal Ubuntu / CPU72.1%70.3%–74.0%0.701—
Quark-LLocal Ubuntu / CPU70.3%68.0%–72.5%0.686—
Quark-HLocal Ubuntu / CPU72.9%71.5%–74.2%0.716—
Quark-G†Local Ubuntu / CUDA67.3%65.6%–68.9%0.670—
LaBraM at Scale · Atom-L and Molecule-L
Atom-LLocal Ubuntu / CPU72.7%70.6%–74.4%0.710—
Molecule-LLocal Ubuntu / CPU71.9%70.2%–73.4%0.697—
REVE Small · early checkpoint (epoch 15 of 200)
REVE SmallLocal Ubuntu / CPU64.5%62.5%–66.6%0.626—

† Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page).

Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04).

The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “Does pretraining help EEG foundation models like LaBraM and CBraMod?”.

Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed.

Limits

Balanced quality-screened scalp subset, not whole-night deployment prevalence and not ear-EEG. Frozen encoders average fifteen 2-second representations; EEGNet receives 10 epochs.

The protocol, step by step

  1. 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once.
  2. Retain six named scalp electrodes; exclude mastoids. Drop any epoch with source per-channel missing-value flag in those six electrodes, actual nonfinite values, any channel std<0.01uV, or peak-to-peak>1000uV. Use full30s at200Hz, no further filtering/reference/amplitude transformation. Select up to30 evenly spaced eligible epochs per participant/class across available nights.
  3. One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 10 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100).
  4. Labels: Wake, N1, N2, N3, REM.
  5. Uniform-guessing reference: 20.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap.
  6. Lightweight balanced quality-screened subset; scores do not describe natural sleep-stage prevalence or the entire73780epoch release.
  7. Two original corrupt sessions and boundary epochs were excluded by the uploader.
  8. 200Hz data may have finite replacements despite original missing-value flags; source flags are therefore enforced instead of relying on finite checks alone.
  9. One prediction per30s epoch. Frozen encoders pool15nonoverlapping2s segments; no sequence context across epochs.
  10. All nights from a participant remain together; pretraining overlap unknown.
  11. Quality thresholds fixed before any scores; source artifacts may remain.

Stability

One fixed seed and training budget; multi-seed sensitivity pending.

Pretraining exposure

Unknown unless explicitly documented; no unseen-pretraining claim.

Checked 2026-10-04. LaBraM: not in the authors' published pretraining list (checked 2026-10-04) source ↗; CBraMod: not in the authors' published pretraining list (checked 2026-10-04) source ↗. The released sentence above predates this check. Further encoders: the exposure column of their table above → · EEGPT: its own section above →

Notes on the methods

Model terms

Source and licence

Permitted scope Personal noncommercial research; aggregate results only

Privacy Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. According to the 2025 data descriptor, the informed consent form did not mention publication, and before release the GDPR office of Region Midt judged the data fully anonymised: consent covered the study, and the public release rests on that anonymisation judgement. The full review note is in the protocol JSON ↓

Public-data register note Clear upstream CC0 and study ethics evidence; use only the pinned scalp-signal derivative and publish aggregates, with provenance disclosed. Consent covered the study, not publication; the public release rests on a GDPR anonymization assessment (amended 2026-09-22).

Rights reviewed · against doi.org ↗ · www.nature.com ↗ · doi.org ↗

Reproducibility record

Protocol id
parallel-fixed-subject-folds-v1/eesm19-scalp-sleep
Dataset release
Hugging Face processed mirror of OpenNeuro ds005185 1.0.2 · mirror fb045012b58a0b755f9fedf5931dff9b64a9797f · upstream 0857858f7a2ba1582930f23eca3ec56f90a96da9
Hardware
Local Ubuntu / CUDA
Scoring stage sum
114.7 s · may include accelerator waiting
Audit record SHA-256
e6cbf1b199bf2902b9c46e4575ebfa8aff47952625df3d71a1d720b11cc27291
Summary SHA-256
b7c784c17d6b33c16ffca5df42d3fd8f3d7bf14960675a7278089c4903886567
Protocol SHA-256
dace8a98e480ae7ee890dd6f498d5753a90e1573813b6e174f3c14bcbae8a8ff

Rows of core-matrix-update-20261106

ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe

Protocol SHA-256
a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c
Audit record SHA-256
f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e
Hardware
Local Ubuntu / CUDA · Local Ubuntu / CPU
Row intervals
participant bootstrap, 10,000 draws
Paired-difference intervals
paired participant bootstrap, 10,000 draws

One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs.

Downloads

This protocol's rows in core-matrix-update-20261106 · CSV ↓ · Core matrix: rows run under its exact protocols · JSON ↓

Other protocols

All protocols → · The core matrix on the home page → · All datasets → · All methods →

Cite this page

BCI Report (2026). Sleep staging on EESM19 scalp subset. https://bci.report/protocols/sleep-scalp/

Figures from releases core-matrix-update-20261106 (2026-11-06), logpower-ridge-update-20261106 (2026-11-06), foundation-checkpoints-update-20261106 (2026-11-06), eegpt-core-update-20261013 (2026-10-13), foundation-models-update-20261004 (2026-10-04), research-preview-20260920 (2026-09-20). Cite the upstream dataset as well: its credit is on this page.

Every release is archived on Zenodo: doi:10.5281/zenodo.23123296. BibTeX for the site and its releases → · CITATION.cff ↗