Twu31/bci-report

BCI Report

Every EEG decoding score, reported with the protocol that produced it.

Website · 中文 · Code · Data use & privacy · Every dataset · Every method · Data API

121 reviewed measurements from public EEG datasets, each carrying the cohort, electrode count, evaluation mode, chance level and training budget that produced it. Core release research-preview-20260920, reviewed 2026-09-20; later batches below, each reviewed on its own.

from datasets import load_dataset

load_dataset("Twu31/bci-report", "results")   # 39 protocol × model scores
load_dataset("Twu31/bci-report", "topics")    # 82 deployment-condition measurements

What is in here

Two separate bodies of measurement. Not one ranking, and rows from one do not belong in a table with rows from the other.

Config	Rows	What it covers	Unit
results	39	9 methods × 8 fixed protocols, 7 public datasets	percent
topics	82	4 deployment questions: dry-vs-wet, on-the-move, calibration-budget, does-pretraining-help	proportion [0,1]
contrasts	18	paired within-participant differences	percentage points
seed_sensitivity	5	repeated training runs of one model	percent
foundation_models	128	16 further foundation-model checkpoints × 8 protocols, frozen probes	percent

Alongside: protocols/ (full descriptor per protocol — preprocessing, split, budget, audit hashes), snapshot.json and deployment-topics.json (the reviewed exports the website itself reads).

Every row in results.csv carries its own license, license_url and attribution, so a row lifted out of that table keeps its credit with it; topics.csv cites its sources in deployment-topics.json under dataset_citations.

evidence-update.json — the 22 September 2026 batch, reviewed under its own manifest: a paired in-ear versus scalp sleep comparison (EESM23, 10 people, identical epochs), four posterior electrodes against all sixteen for eyes open or closed (Alpha Waves, 19 of the source's 20 recordings), a physical-phantom artifact test reporting correlation and predictive R² (dimensionless; R² is negative where the decoder fails and is published as measured), and the adaptation roadmap as it stood that day (planned; the measured results are in adaptation-update.json, below). Heterogeneous by design, so it ships as JSON rather than as a table.

clinical-update.json — the 23 September 2026 batch, reviewed under its own manifest: a 149-person case/control comparison on a Parkinson's disease data set (100 patients, 49 controls, one site) published beside an age-and-sex-only confound comparator, plus three holds that produced no score. Research results, not diagnosis: not diagnostic accuracy, not clinical validation, and no interpretation of any individual. No participant rows, no clinical scores and no per-group demographics are published.

context-update.json — the 27 September 2026 batch: a P300 calibration carried from a PC screen to a VR headset and back (21 people, within-person, two fixed baselines, two timing schemes), and walking-speed classification from dry-electrode EEG printed beside a movement-nuisance comparator (58 people), and a four-person pilot of SSVEP windows accepted when no command was intended — window-level counts, always with coverage beside accuracy. Fixed CPU baselines; no foundation-model or fine-tuning result. One one-person pilot is published as status only.

adaptation-update.json — reviewed 1 October 2026: LaBraM adapted to new people on EEGMAT mental arithmetic (36 people, five participant-disjoint folds) three ways — training only a classification head, the last block as well, or rank-4 LoRA — with the same checkpoint, folds, starting heads, batch order and five-epoch recipe, over three seeds. Cohort means, paired changes with people helped and harmed, per-seed means, trainable parameters and training time. The head-only arm is a short gradient-trained head, so the file points to the core matrix's frozen LaBraM ridge readout on the same folds for scale. One fixed recipe, not a tuned ranking; no memory figures. A next-day experiment on a source under editorial hold is listed as status only, with no figure.

extension-update.json — reviewed 2 October 2026, two fixed classical baselines. The asynchronous SSVEP non-control test extended to the twenty other people of the same release, scored under two rejection rules fixed before scoring: a threshold fitted on the four-person pilot and a personal threshold fitted on 96 of each person's own windows. Detection balanced accuracy, coverage, correct-and-accepted output and false acceptance per non-control state for both, with people helped, harmed and tied. A better detector here is not better command accuracy, and these are window rates, not false activations per hour. And LTRSVP (PhysioNet, doi:10.13026/C2KX0P; Matran-Fernandez and Poli, PLoS ONE 2017, doi:10.1371/journal.pone.0178498; ODC-By 1.0): a P300 target decoder trained on one recording at 5 or 10 Hz and tested on a different recording of the same person at 10 Hz, nine people, with the full 3×3 rate matrix. Within a rate, run b followed run a after a long break; across rates the original study presented the rates from the lowest to the highest, not randomised, and how the released files map onto that sequence is not documented, so a cross-rate cell also differs in elapsed time, fatigue and practice. The paired interval crosses zero, and rate and recording change together: not a causal effect of image rate.

large-source-update.json — reviewed 3 October 2026, two new sources and two separate questions, each with two fixed classical CPU baselines and no shared ranking. Dreem Open Datasets (Guillot et al., IEEE TNSRE 2020, doi:10.1109/TNSRE.2020.3011181; deposit doi:10.5281/zenodo.15900394, MIT as the deposit declares it): five-stage sleep staging against the publisher's consensus in 25 healthy sleepers (DOD-H) and 55 people with obstructive sleep apnoea (DOD-O), two separate experiments, never compared with each other. A training prior that always predicts N2 and a spectral ridge, with accuracy, balanced accuracy, macro F1 and Cohen's kappa, the paired balanced-accuracy gain, and the ridge's per-stage recall, precision and F1. Read accuracy beside balanced accuracy: the ridge reaches about 72% accuracy but 49–57% balanced accuracy and never predicts N1, so its N1 precision is null — not defined, not zero. The prior's balanced accuracy sits slightly above one fifth: a floor, not a chance level. Not clinical, not diagnosis; the recordings' physical units are unresolved, so amplitude-sensitive models are held with no figure. And OpenBMI (Lee et al., GigaScience 2019, doi:10.1093/gigascience/giz002; data doi:10.5524/100542, CC0 1.0): a motor-imagery decoder trained on each person's first session and tested on the final 60 trials of their second, after 0, 10, 20 or 40 labelled trials from it, 51 people. Every mean change carries how many people declined; the relative-PSD gain at 40 trials has an interval that includes zero and is not established. An expanded cohort under the same method, not an independent replication.

reliable-decisions-update.json — reviewed 4 October 2026: route 1 of the decision-research roadmap, when a decoder should decline to decide. No classifier was retrained: the saved test scores of models already published here — spectral ridge, EEGNet and two LaBraM arms on EEGMAT (36 people), standard CCA, CBraMod and EEGNet on BETA (70 people) — rescored under a fixed threshold on calibrated confidence, a coverage target, a certified selective risk (SGR) and a learned reject option, with NLL, ECE and Brier score and a person-specific recalibration whose label cost per new person is counted. Every full-coverage accuracy equals the published one. A fixed threshold accepted 28.4% of BETA trials for one method and 1.5% for another, so compare methods at matched coverage, not at a shared cut-off; how many people had nothing accepted travels with every coverage. The learned reject option erred less than the calibrated confidence for no method. The certified risk accepted nothing on EEGMAT — no error rate, not a zero one — and on BETA some certified folds exceeded their own target: a nominal guarantee across people. Calibration data are participant-disjoint inner cross-fit scores, and rejecting likely errors on known classes is not detecting unfamiliar input. Every contrast carries its paired person-bootstrap interval and a verdict; nothing is ranked where intervals overlap. Every protocol is balanced or uniform by design, so these are method comparisons, never deployment error rates. ds003810 is a crude secondary protocol; the idle protocol (another unit) and a source under editorial hold are left out.

foundation-models-update.json and foundation-models.csv — reviewed 4 October 2026: eleven further EEG foundation models, sixteen encoder checkpoints (REVE Base and Large, LUNA Base and Large, BrainOmni Base, CodeBrain, EEGMamba, ST-EEGFormer Base and Large, four eeg-fm-masking checkpoints, ERP-FM Base, SingLEM and ZUNA 1.1), run as frozen probes on the same eight protocols with the published LaBraM and CBraMod recipe and only the encoder swapped; nine of them also adapted on EEGMAT with the 1 October recipe (a head trained on the frozen encoder against rank-4 LoRA, three seeds; the adaptation rows are in the JSON). foundation-models.csv stacks the site's eight per-protocol CSVs: its first columns are the results config's, in percent, and the rest give each row's status, panel (matrix or masking ablation), chance flag, pretraining exposure, licence and footnote. Kept apart from results, and not a ranking: rows are grouped by family, and one is above or below another only where their 95% intervals do not overlap. Sleep staging is the only protocol where new cells lie entirely above every published row; elsewhere no new cell lies above the published non-foundation row with the highest point estimate, and on BETA no new row lies above standard CCA, which needs no training. Two BrainOmni cells were not run (its tokenizer needs two-second windows): empty, with the reason, never zero. Pretraining exposure is a sourced statement: BETA is in ST-EEGFormer's published pretraining list and TMNRED in SingLEM's; ZUNA 1.1 lists no pretraining data, so its cells are unknown; every other cell, LaBraM's and CBraMod's included, is not in the authors' published pretraining list as checked on 4 October 2026 — not proof that a recording was never seen. Weights licences travel with every row (REVE Responsible Use License, LUNA CC BY-ND 4.0, ERP-FM CC BY-NC-SA 4.0; ZUNA 1.1's model card: research use only, not for diagnosis or clinical use). Timings came from a shared GPU and are not published.

9 of 18 catalogued methods have been scored. A method with no row has not been run, which is not the same as having failed.

What this does not contain

No raw EEG, no per-participant scores, no participant identifiers, no recordings, no embeddings, no model weights. This is a results table, not a data mirror. Reach the recordings through the source column and follow that source's own terms — several of the underlying datasets are redistributable and several are not, and this repository does not relicense any of them.

Read this before ranking anything

• Chance level differs per protocol. 50% for the binary tasks, 2.5% for 40-class BETA SSVEP, 20% for five-stage sleep. A 57% SSVEP result is far above chance; a 57% binary result is barely above it. Sorting the table by primary_percent across protocols compares numbers that do not share a scale.
• Intervals are descriptive, not confidence intervals. They are bootstrap spreads over the scoring set. They do not support significance claims.
• Most comparisons use one seed. Two protocols add a three-seed EEGNet sensitivity check; the main table keeps its original fixed seed.
• Electrode subsets are not headsets. Four- and eight-electrode subsets of laboratory recordings do not validate a physical four- or eight-channel device.
• Pretraining overlap is checked against published lists, not proven absent. For LaBraM and CBraMod, none of the seven datasets is in the authors' published pretraining list (checked 4 October 2026; every source is in foundation-models-update.json). The released protocol files keep their original sentence, that overlap is unknown unless documented; recording-level overlap was not audited, so a frozen encoder may still have seen related data.

And for topics.csv specifically:

• The unit changes. value is a proportion in [0,1] here, a percent in results.csv. contrasts.csv differences are percentage points. Concatenating the two tables without rescaling silently divides one of them by 100.
• 82 measurements are not 82 studies. They are protocol-specific rows across 4 questions; a question's rows share a cohort and a protocol, so they are not independent evidence.
• Seed spread is not a confidence interval. seed-sensitivity.csv reports what repeated training runs of the same model on the same data did. It describes the optimizer, not the population, and it cannot be read as an error bar on a participant mean.
• These four questions do not extend the 8-protocol matrix. Different protocols, different cohorts, separately reviewed. Keep them apart.

A small cohort is close to its parts

The idle protocol has a cohort of four, and at that size a published mean can be turned back into a count: 1.7% of 60 command trials is one detection, and since two of the four participants selected an always-abstain threshold, that detection belongs to one of the remaining two.

These counts are published anyway, because a false-activation rate of zero produced by people who never activate is the more misleading number. What a reader cannot recover is which person is which — participants are unidentified here and in the public source recordings, and no demographic, session or ordering information is published that would let anyone line them up.

Treat "aggregate" as a description of what is published, not as a claim that it cannot be inverted.

Protocols

Protocol	Dataset	Cohort	Electrodes	Chance	Metric
mi-rest	ds003810	10	15	50.0%	Balanced accuracy
idle	ds005342	4	17	none — see note	Command detection ≤3s
beta-8ch	BETA	70	8	2.5%	Balanced accuracy
beta-4ch	BETA	70	4	2.5%	Balanced accuracy
arithmetic-rest	EEGMAT	36	19	50.0%	Balanced accuracy
p300-target	ds006593	21	19	50.0%	Balanced accuracy
semantic-target	TMNRED / ds005383	30	30	50.0%	Balanced accuracy
sleep-scalp	EESM19 scalp subset	20	6	20.0%	Balanced accuracy

Source datasets and attribution

Results were computed from these public datasets. Credit belongs to their original authors; this repository adds only the measurements.

Dataset	License	Attribution
ds003810	CC0-1.0	Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/
EEGMAT	Open Data Commons Attribution License 1.0	Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z.
EESM19 scalp subset	CC0-1.0 declared by upstream and mirror	Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed.
BETA	CC BY 4.0	Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA.
TMNRED / ds005383	OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0	Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0.
ds006593	CC0-1.0	OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source.
ds005342	CC0-1.0	OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source.

Corrections to released wording

Released files are fixed bytes, so wording errors are corrected here and on the site rather than in place (full list: https://bci.report/releases/#corrections).

• deployment-topics.json calls the original ensemble-TRCA method a three-filter-bank experiment. The original paper and the reference code's tutorials use five sub-bands; three is only the default of the code's functions. No score changes.
• For ds003810, the OpenNeuro record asks users to cite Peterson, Galván, Hernández and Spies, A feasibility study of a complete low-cost consumer-grade brain-computer interface system, Heliyon 6(3):e03425 (2020), doi:10.1016/j.heliyon.2020.e03425 — in addition to the Data in Brief description linked in the rows.

Citation

Cite BCI Report and the release you used — this card was built at release foundation-models-update-20261004 (2026-10-04) — and the upstream dataset each figure was computed on: the attribution table above, and every dataset page at https://bci.report/datasets/, gives its credit. The same citation is in CITATION.cff, which GitHub's "Cite this repository" reads, and every topic, dataset and method page on the site ends with a "Cite this page" block naming the releases its figures come from. Every release is archived on Zenodo; the concept DOI 10.5281/zenodo.23123296 resolves to the newest one.

@misc{bcireport,
  title        = {BCI Report: public EEG decoding results reported with their protocol},
  author       = {{BCI Report}},
  year         = {2026},
  howpublished = {\url{https://bci.report}},
  doi          = {10.5281/zenodo.23123296},
  note         = {Release foundation-models-update-20261004}
}

License

The cc-by-4.0 tag covers the aggregate result tables and protocol descriptors in this repository — measurements this project produced. It does not and cannot relicense the underlying recordings, whose terms are listed per row and per dataset above. The code that produced them, on GitHub, is under the MIT License.

Corrections

Scientific corrections, benchmark-method questions, and rights, privacy or attribution concerns are welcome, including ones that would withdraw a published result — affected results can be withheld while an issue is reviewed. Open a discussion here, or write to contact@bci.report for scientific corrections and privacy@bci.report for rights, privacy and withdrawal requests. Please do not send raw EEG, participant names or health records.

Full source review, privacy reasoning and interpretation caveats: https://bci.report/data-use/