EEGMAT: EEG decoding results

Name: BCI Report: results on EEGMAT
Creator: BCI Report
License: https://creativecommons.org/licenses/by/4.0/

Mental arithmetic / rest

EEGMAT: EEG decoding results

EEGMAT (EEG During Mental Arithmetic Tasks) is a PhysioNet database, version 1.0.0 published December 2018, with two 60-second artifact-free EEG recordings from each of 36 participants: one before and one during a serial-subtraction task. Contributed from Igor Sikorsky Kyiv Polytechnic Institute and described by Zyma et al. in Data (2019).

Also known as EEGMAT · EEG During Mental Arithmetic Tasks · Electroencephalograms during Mental Arithmetic Task Performance · Zyma et al. (2019)

Description sources physionet.org · doi.org

Where it appears

Published results

Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap.

Core matrix · arithmetic & rest

Source page: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · experiments.json, core-matrix-20261106-arithmetic-rest.csv

Spectral ridge

56.8% (54.6%–58.9%)

LaBraM

64.6% (60.5%–68.6%)

CBraMod

62.3% (58.1%–66.5%)

EEGNet

67.6% (62.8%–72.5%)

TS+LR · OAS · 4–40 Hz

65.5% (59.6%–71.8%)

MDM · OAS · 4–40 Hz

71.5% (65.5%–77.8%)

ShallowFBCSPNet

67.9% (63.4%–72.4%)

Deep4Net

61.3% (57.5%–65.2%)

CSP+LDA

68.3% (62.0%–74.3%)

Balanced accuracy, 9 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
Spectral ridge	Serial subtraction versus resting EEG	Balanced accuracy	56.8%54.6%–58.9%	36
LaBraM	Serial subtraction versus resting EEG	Balanced accuracy	64.6%60.5%–68.6%	36
CBraMod	Serial subtraction versus resting EEG	Balanced accuracy	62.3%58.1%–66.5%	36
EEGNet	Serial subtraction versus resting EEG	Balanced accuracy	67.6%62.8%–72.5%	36
TS+LR · OAS · 4–40 Hz	Serial subtraction versus resting EEG	Balanced accuracy	65.5%59.6%–71.8%Primary row · one fixed configuration (MOABB-reference settings)	36
MDM · OAS · 4–40 Hz	Serial subtraction versus resting EEG	Balanced accuracy	71.5%65.5%–77.8%Secondary row · one fixed configuration (MOABB-reference settings)	36
ShallowFBCSPNet	Serial subtraction versus resting EEG	Balanced accuracy	67.9%63.4%–72.4%One fixed seed and training budget, as the EEGNet cell	36
Deep4Net	Serial subtraction versus resting EEG	Balanced accuracy	61.3%57.5%–65.2%One fixed seed and training budget, as the EEGNet cell	36
CSP+LDA	Serial subtraction versus resting EEG	Balanced accuracy	68.3%62.0%–74.3%Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP.	36

Further foundation encoders, frozen · arithmetic & rest

Source page: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · foundation-models-arithmetic-rest.csv

REVE Base · Frozen encoder + ridge head

68.8% (64.1%–73.2%)

REVE Large · Frozen encoder + ridge head

69.0% (63.3%–74.7%)

LUNA Base · Frozen encoder + ridge head

60.7% (56.9%–64.4%)

LUNA Large · Frozen encoder + ridge head

57.4% (54.3%–60.6%)

BrainOmni Base · Frozen encoder + ridge head

61.6% (58.4%–64.9%)

CodeBrain · Frozen encoder + ridge head

63.0% (59.5%–66.5%)

EEGMamba · Frozen encoder + ridge head

60.3% (57.0%–63.4%)

ST-EEGFormer Base · Frozen encoder + ridge head

64.5% (59.0%–70.0%)

ST-EEGFormer Large · Frozen encoder + ridge head

63.4% (57.8%–69.0%)

eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head

63.6% (59.1%–68.1%)

ERP-FM Base · Frozen encoder + ridge head

61.1% (57.3%–65.1%)

SingLEM · Frozen encoder + ridge head

51.5% (49.2%–53.8%)

ZUNA 1.1 · Frozen encoder + ridge head

53.3% (49.5%–57.1%)

eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation

61.6% (56.4%–66.5%)

eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation

64.3% (60.2%–68.2%)

eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation

62.5% (58.5%–66.2%)

Balanced accuracy, 16 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
REVE Base	Frozen encoder + ridge head	Balanced accuracy	68.8%64.1%–73.2%	36
REVE Large	Frozen encoder + ridge head	Balanced accuracy	69.0%63.3%–74.7%	36
LUNA Base	Frozen encoder + ridge head	Balanced accuracy	60.7%56.9%–64.4%	36
LUNA Large	Frozen encoder + ridge head	Balanced accuracy	57.4%54.3%–60.6%	36
BrainOmni Base	Frozen encoder + ridge head	Balanced accuracy	61.6%58.4%–64.9%	36
CodeBrain	Frozen encoder + ridge head	Balanced accuracy	63.0%59.5%–66.5%	36
EEGMamba	Frozen encoder + ridge head	Balanced accuracy	60.3%57.0%–63.4%	36
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	64.5%59.0%–70.0%	36
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	63.4%57.8%–69.0%	36
eeg-fm-masking MAE (r=9 cm, L=2)	Frozen encoder + ridge head	Balanced accuracy	63.6%59.1%–68.1%	36
ERP-FM Base	Frozen encoder + ridge head	Balanced accuracy	61.1%57.3%–65.1%	36
SingLEM	Frozen encoder + ridge head	Balanced accuracy	51.5%49.2%–53.8%	36
ZUNA 1.1	Frozen encoder + ridge head	Balanced accuracy	53.3%49.5%–57.1%Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use.	36
eeg-fm-masking JEPA (r=9 cm, L=2)	Frozen encoder + ridge head · masking ablation	Balanced accuracy	61.6%56.4%–66.5%	36
eeg-fm-masking MAE (r=one channel, L=1)	Frozen encoder + ridge head · masking ablation	Balanced accuracy	64.3%60.2%–68.2%	36
eeg-fm-masking JEPA (r=one channel, L=1)	Frozen encoder + ridge head · masking ablation	Balanced accuracy	62.5%58.5%–66.2%	36

Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. Every weights licence and its terms →

EEGPT, frozen encoder · arithmetic & rest

Source page: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · eegpt-core-update.json

Method	Condition	Metric	Value	People
EEGPT	Frozen encoder + ridge head	Balanced accuracy	65.7%60.9%–70.3%	36

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.

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.

EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. Its licence and terms →

Pretraining · fixed readout

Source page: Does pretraining help? · deployment-topics.json, eegpt-core-update.json

Method	Condition	Metric	Value	People
LaBraM	Mental workload · pretrained	Balanced accuracy	64.6%	36
LaBraM	Mental workload · random initialization (mean)	Balanced accuracy	56.5%	36
CBraMod	Mental workload · pretrained	Balanced accuracy	62.3%	36
CBraMod	Mental workload · random initialization (mean)	Balanced accuracy	56.4%	36
EEGPT	Mental workload · pretrained	Balanced accuracy	65.7%	36
EEGPT	Mental workload · random initialization (mean)	Balanced accuracy	59.2%	36

Pretraining · train-selected readout

Source page: Does pretraining help? · deployment-topics.json, eegpt-core-update.json

Method	Condition	Metric	Value	People
LaBraM	Mental workload · pretrained	Balanced accuracy	64.4%	36
LaBraM	Mental workload · random initialization (mean)	Balanced accuracy	56.5%	36
CBraMod	Mental workload · pretrained	Balanced accuracy	68.4%	36
CBraMod	Mental workload · random initialization (mean)	Balanced accuracy	61.7%	36
EEGPT	Mental workload · pretrained	Balanced accuracy	66.0%	36
EEGPT	Mental workload · random initialization (mean)	Balanced accuracy	58.6%	36

Pretraining · fixed classical control

Source page: Does pretraining help? · deployment-topics.json

Method	Condition	Metric	Value	People
Log-covariance ridge	Mental workload	Balanced accuracy	61.9%56.2%–67.6%	36

Model adaptation · LaBraM on new people, same task, zero labels from the test person

Source page: Which part to update? · Chance level 50.0% · adaptation-update.json

LaBraM · Head only · encoder frozen

56.6% (54.2%–59.0%)

LaBraM · Last block + head

65.7% (61.6%–69.7%)

LaBraM · LoRA rank 4 + head

64.1% (60.2%–67.8%)

Balanced accuracy, 3 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
LaBraM	Head only · encoder frozen	Balanced accuracy	56.6%54.2%–59.0%	36
LaBraM	Last block + head	Balanced accuracy	65.7%61.6%–69.7%	36
LaBraM	LoRA rank 4 + head	Balanced accuracy	64.1%60.2%–67.8%	36
LaBraM	LoRA rank 4 + head minus head only	Paired difference	+7.5 pp+4.8 pp–+10.1 pp	36
LaBraM	Last block + head minus head only	Paired difference	+9.1 pp+6.0 pp–+12.2 pp	36
LaBraM	LoRA rank 4 + head minus last block + head	Paired difference	−1.6 pp−4.4 pp–+1.0 pp	36

Model adaptation, further encoders · new people, same task, one fixed recipe

Source page: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · foundation-models-update.json

Method	Condition	Metric	Value	People
REVE Base	Frozen encoder + trained head	Balanced accuracy	64.0%60.4%–67.7%1,026 trainable parameters.	36
REVE Base	LoRA rank 4 + head	Balanced accuracy	66.3%61.9%–70.7%181,250 trainable parameters.	36
REVE Base	LoRA minus frozen + head, same people and folds	Paired difference	+2.3 pp+0.1 pp–+4.6 pp21 people improved, 13 got worse, 2 unchanged. The interval excludes zero.	36
REVE Large	Frozen encoder + trained head	Balanced accuracy	65.6%61.1%–70.1%2,434 trainable parameters.	36
REVE Large	LoRA rank 4 + head	Balanced accuracy	69.2%63.4%–74.9%430,466 trainable parameters.	36
REVE Large	LoRA minus frozen + head, same people and folds	Paired difference	+3.6 pp+0.5 pp–+6.8 pp21 people improved, 13 got worse, 2 unchanged. The interval excludes zero.	36
LUNA Base	Frozen encoder + trained head	Balanced accuracy	58.2%54.5%–61.7%514 trainable parameters.	36
LUNA Base	LoRA rank 4 + head	Balanced accuracy	60.0%55.9%–63.9%37,378 trainable parameters.	36
LUNA Base	LoRA minus frozen + head, same people and folds	Paired difference	+1.8 pp+0.1 pp–+3.3 pp23 people improved, 11 got worse, 2 unchanged. The interval excludes zero.	36
BrainOmni Base	Frozen encoder + trained head	Balanced accuracy	64.5%59.6%–69.2%16,386 trainable parameters.	36
BrainOmni Base	LoRA rank 4 + head	Balanced accuracy	64.8%59.6%–69.7%106,498 trainable parameters.	36
BrainOmni Base	LoRA minus frozen + head, same people and folds	Paired difference	+0.3 pp−1.4 pp–+1.8 pp18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established.	36
CodeBrain	Frozen encoder + trained head	Balanced accuracy	66.5%62.3%–70.6%7,602 trainable parameters.	36
CodeBrain	LoRA rank 4 + head	Balanced accuracy	67.1%62.6%–71.6%58,802 trainable parameters.	36
CodeBrain	LoRA minus frozen + head, same people and folds	Paired difference	+0.6 pp−0.7 pp–+1.9 pp18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established.	36
EEGMamba	Frozen encoder + trained head	Balanced accuracy	60.4%55.7%–65.2%7,602 trainable parameters.	36
EEGMamba	LoRA rank 4 + head	Balanced accuracy	60.8%56.3%–65.3%90,930 trainable parameters.	36
EEGMamba	LoRA minus frozen + head, same people and folds	Paired difference	+0.4 pp−0.7 pp–+1.5 pp18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established.	36
ST-EEGFormer Base	Frozen encoder + trained head	Balanced accuracy	63.2%59.7%–66.7%1,538 trainable parameters.	36
ST-EEGFormer Base	LoRA rank 4 + head	Balanced accuracy	65.1%61.0%–69.4%148,994 trainable parameters.	36
ST-EEGFormer Base	LoRA minus frozen + head, same people and folds	Paired difference	+1.9 pp−0.2 pp–+4.1 pp20 people improved, 16 got worse, 0 unchanged. The interval includes zero: no change is established.	36
SingLEM	Frozen encoder + trained head	Balanced accuracy	50.9%49.5%–52.3%1,218 trainable parameters.	36
SingLEM	LoRA rank 4 + head	Balanced accuracy	60.6%56.7%–64.4%33,986 trainable parameters.	36
SingLEM	LoRA minus frozen + head, same people and folds	Paired difference	+9.7 pp+6.1 pp–+13.3 pp30 people improved, 6 got worse, 0 unchanged. The interval excludes zero.	36
ZUNA 1.1	Frozen encoder + trained head	Balanced accuracy	51.2%49.8%–52.7%1,218 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use.	36
ZUNA 1.1	LoRA rank 4 + head	Balanced accuracy	56.9%53.7%–60.1%296,130 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use.	36
ZUNA 1.1	LoRA minus frozen + head, same people and folds	Paired difference	+5.7 pp+2.5 pp–+8.7 pp28 people improved, 7 got worse, 1 unchanged. The interval excludes zero. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use.	36

Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0. No model’s authors endorse these results. Every weights licence and its terms →

Reliable decisions · when to decline a decision

Source page: When not to act · reliable-decisions-update.json

Method	Condition	Metric	Value	People
Spectral ridge	Fixed threshold: calibrated confidence at least 0.8	Windows accepted	3 / 2,16033 of 36 people had nothing accepted.	36
Spectral ridge	Learned reject option minus calibrated confidence, same classifier	Error among the 80% most certain, difference	+0.5 pp−1.2 pp–+1.7 ppNo difference resolved: the interval includes zero.	36
EEGNet	Fixed threshold: calibrated confidence at least 0.8	Windows accepted	126 / 2,16019 of 36 people had nothing accepted.	36
EEGNet	Learned reject option minus calibrated confidence, same classifier	Error among the 80% most certain, difference	+2.3 pp+0.6 pp–+4.1 ppDifference resolved: the interval excludes zero.	36
LaBraM · head only, encoder frozen	Fixed threshold: calibrated confidence at least 0.8	Windows accepted	1 / 2,16035 of 36 people had nothing accepted.	36
LaBraM · head only, encoder frozen	Learned reject option minus calibrated confidence, same classifier	Error among the 80% most certain, difference	+0.6 pp−0.5 pp–+2.7 ppNo difference resolved: the interval includes zero.	36
LaBraM · LoRA rank 4 + head	Fixed threshold: calibrated confidence at least 0.8	Windows accepted	44 / 2,16024 of 36 people had nothing accepted.	36
LaBraM · LoRA rank 4 + head	Learned reject option minus calibrated confidence, same classifier	Error among the 80% most certain, difference	+1.9 pp+0.6 pp–+3.5 ppDifference resolved: the interval excludes zero.	36

Limits: Balanced by design. Error rates describe a 1:1 mix and are a method comparison, not a deployment rate. Class equals recording, so recording-level differences are part of what any model or selector can use. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class (https://arxiv.org/abs/2603.13324). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. Every limitation of route 1 →

Source and licence

EEGMAT

Credit 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.

Open Data Commons Attribution License 1.0 ↗ · Dataset record ↗

BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’.

Row in the open register →

All datasets → · All methods →

Cite this page

BCI Report (2026). EEGMAT: EEG decoding results. https://bci.report/datasets/eegmat/

Figures from releases core-matrix-update-20261106 (2026-11-06), eegpt-core-update-20261013 (2026-10-13), foundation-models-update-20261004 (2026-10-04), reliable-decisions-update-20261004 (2026-10-04), adaptation-update-20261001 (2026-10-01), deployment-topics-20260920-v1 (2026-09-20), 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 ↗