# How to read EEG decoding results

Start here

A five-minute field guide to scalp EEG tasks, fixed protocols, chance levels, and descriptive intervals — paired with six interactive checks drawn straight from published runs.

- 01What EEG decoding checks
- 02Fixed protocols
- 03Chance & 95% intervals
- 04Six data checks
- 05Why no overall score
- 06Where to go next

01 · Signals & scope

## What EEG decoding is, and what this site checks

Tap any dotted term — such as [EEG](https://bci.report/glossary/#eeg), [SSVEP](https://bci.report/glossary/#ssvep), or [P300 / ERP](https://bci.report/glossary/#p300-erp) — for an inline definition.

Electroencephalography (EEG) records tiny electrical voltage changes across sensors placed on the scalp. A brain–computer interface (BCI) turns those noisy waveforms into discrete selections or state estimates — such as choosing a flickering target on a screen, detecting a rare flash, distinguishing imagined hand movement from rest, or scoring sleep stages.

BCI Report runs classical signal-processing methods, compact neural networks, and frozen or adapted foundation encoders offline on public datasets under audited recipes. Every figure is aggregate cohort-level evidence: this site never performs live EEG inference, never publishes individual-person scores, and makes no clinical or diagnostic claim.

Visual frequency

### SSVEP

Steady-state visual evoked potentials track flickering visual targets at distinct frequencies over posterior scalp sites.

40 targets · 2.5% chance

Time-locked wave

### P300 / ERP

Event-related potentials are brief voltage deflections ~300 ms after a rare or attended visual or semantic target.

2 classes · 50.0% chance

Sensorimotor rhythm

### Motor imagery

Imagining a limb movement without moving shifts mu and beta power over motor cortex, though patterns vary widely across people.

2 classes · 50.0% chance

Overnight frames

### Sleep staging

Scoring 30-second overnight windows into five sleep stages (Wake, N1, N2, N3, REM) where rare stages matter as much as common ones.

5 stages · 20.0% chance

02 · Comparability

## How a fixed protocol works

See how a [participant-disjoint split](https://bci.report/glossary/#participant-disjoint) and locked recipe keep method comparisons clean.

Two papers both labeled “motor imagery” or “sleep staging” are rarely interchangeable. Changing the dataset, electrode layout, window length, or split rule can swing accuracy by twenty percentage points even for the exact same code.

Every core benchmark table on BCI Report therefore locks four ingredients into a fixed protocol before any method runs:

01

### Cohort

Exact dataset version and participant count — from a 4-person idle pilot to 70 people on BETA SSVEP.

02

### Electrode montage

Fixed channel set and scalp layout (for instance, 8 vs. 4 posterior electrodes on BETA, or 6 scalp derivations on sleep).

03

### Participant-disjoint split

All recordings from one person stay together on one side of a fold, testing transfer to unseen people rather than memorizing one person’s baseline.

04

### Locked recipe

Identical windowing, preprocessing, training budget, and metric across every method compared inside that column.

03 · Reading a number

## Always read chance level and the 95% interval together

Read [balanced accuracy](https://bci.report/glossary/#balanced-accuracy), [chance level](https://bci.report/glossary/#chance-level), and the [descriptive 95% interval](https://bci.report/glossary/#descriptive-interval) together.

A percentage on its own tells you very little. A 33.7% score on a 40-class SSVEP task sits more than 31 percentage points above its 2.5% chance line, whereas 51.0% on a 2-class P300 task sits right on its 50.0% chance line.

Beside every point estimate, BCI Report prints a descriptive 95% participant-bootstrap interval showing how much scores vary across the held-out people in that cohort. Compare the two real published slices below:

### 40-class SSVEP (BETA · 8 ch · 70 people)

Chance 2.5%

**Published core-matrix rows on beta-8ch.** Dashed line = 2.5% uniform chance; bar = descriptive 95% participant-bootstrap interval.

- CBraMod — Frozen encoder · 95% interval 29.8–37.6%: 33.7%
- EEGNet — Compact model · 95% interval 50.1–61.3%: 55.8%
- Standard CCA — Training-free reference · 95% interval 57.2–69.0%: 63.1%
- FBCCA — Filter-bank reference · 95% interval 75.1–82.8%: 79.1%

**Key reading:**Even the 33.7% row (CBraMod) stands more than 30 percentage points clear of uniform guessing across 40 targets.

### 2-class P300 target ERP (21 people)

Chance 50.0%

**Published core-matrix rows on p300-target.** Dashed line = 50.0% uniform chance; bar = descriptive 95% participant-bootstrap interval.

- LaBraM — Frozen encoder · 95% interval 47.7–51.2%: 49.4%
- Spectral ridge — Band-power baseline · 95% interval 49.6–52.4%: 51.0%
- Temporal ridge — Time-domain baseline · 95% interval 51.9–55.6%: 53.8%
- XDAWNCov+TS+LR — xDAWN + Riemannian · 95% interval 58.8–65.0%: 61.9%

**Key reading:**Spectral ridge prints 51.0% — a larger raw number than 33.7% above — yet its 95% interval (49.6–52.4%) crosses the 50% chance line.

Published values shown in the two illustrative plots above.

| Protocol | Method | Balanced accuracy | Descriptive 95% interval | Chance level |
| --- | --- | --- | --- | --- |
| SSVEP · 8 channels (40 classes) | CBraMod | 33.7% | 29.8–37.6% | 2.5% |
| SSVEP · 8 channels (40 classes) | EEGNet | 55.8% | 50.1–61.3% | 2.5% |
| SSVEP · 8 channels (40 classes) | Standard CCA | 63.1% | 57.2–69.0% | 2.5% |
| SSVEP · 8 channels (40 classes) | FBCCA | 79.1% | 75.1–82.8% | 2.5% |
| P300 target ERP (2 classes) | LaBraM | 49.4% | 47.7–51.2% | 50.0% |
| P300 target ERP (2 classes) | Spectral ridge | 51.0% | 49.6–52.4% | 50.0% |
| P300 target ERP (2 classes) | Temporal ridge | 53.8% | 51.9–55.6% | 50.0% |
| P300 target ERP (2 classes) | XDAWNCov+TS+LR | 61.9% | 58.8–65.0% | 50.0% |

04 · Interactive check

## Guess then reveal: six checks from published results

Before looking at the full matrices, test your intuition on six real findings from our published runs. Pick an answer to reveal the exact figures and protocol link — there is no timer and no score tally.

01 / 06 40-class SSVEP · 70 people

### On 40-class SSVEP with 8 posterior electrodes, does the paired interval of FBCCA minus Standard CCA sit above zero?

- **A.** Yes — the adjusted interval sits cleanly above zero
- **B.** No — the interval crosses zero, so the difference is unresolved
- **C.** No — Standard CCA stays higher than FBCCA

Show published result

FBCCA 79.1% vs. Standard CCA 63.1% · paired mean +16.0 pp (98.75% interval +11.6 to +20.5 pp)

Weighting multiple harmonic sub-bands separates FBCCA from single-band CCA across the 70-person cohort, while neither method requires training data.

[Inspect SSVEP · 8 channels protocol →](https://bci.report/protocols/beta-8ch/)

02 / 06 5-stage sleep · 20 people

### On five-stage sleep staging (chance 20%), does every frozen foundation encoder separate above classical Spectral ridge?

- **A.** Yes — every frozen foundation encoder sits above Spectral ridge
- **B.** No — results split by encoder: some separate above it, others overlap or sit below
- **C.** No — Spectral ridge sits above every frozen encoder

Show published result

Spectral ridge 72.2% (70.7–73.8%) · LaBraM 71.1% (69.0–73.2%) · ST-EEGFormer Large 79.6% (78.2%–81.0%) · EEGPT 75.5%

Foundation models do not act as a single monolith: some checkpoints overlap a fast spectral baseline while others separate above it under the same frozen ridge head.

[Inspect sleep staging protocol →](https://bci.report/protocols/sleep-scalp/)

03 / 06 Motor imagery · 10 people

### On 2-class motor imagery vs. rest with 10 participants, roughly how wide is EEGNet’s descriptive 95% participant-bootstrap interval?

- **A.** Narrow — roughly 1 to 2 percentage points wide
- **B.** Wide — more than 11 percentage points across the 10 people
- **C.** Zero width — a fixed random seed has no interval

Show published result

EEGNet 71.0% · 95% interval 65.0–76.3% (width 11.3 pp across 10 people)

On small cohorts, person-to-person spread spans over ten percentage points, so tiny differences in point estimates cannot be over-interpreted.

[Inspect motor imagery & rest protocol →](https://bci.report/protocols/mi-rest/)

04 / 06 P300 target ERP · 21 people

### On 2-class P300 target ERP (chance 50%), does XDAWNCov+TS+LR separate from Spectral ridge?

- **A.** Yes — time-locked xDAWN filtering separates cleanly above Spectral ridge
- **B.** No — band power captures P300 just as well
- **C.** No — Spectral ridge sits higher than XDAWNCov+TS+LR

Show published result

XDAWNCov+TS+LR 61.9% vs. Spectral ridge 51.0% · paired mean +11.0 pp (98.75% interval +6.8 to +15.1 pp)

P300 is a brief time-locked deflection rather than a steady spectral oscillation; Spectral ridge sits right on the 50% chance line while xDAWN extracts the evoked response.

[Inspect P300 target ERP protocol →](https://bci.report/protocols/p300-target/)

05 / 06 Idle & command · 4-person pilot

### On the cue-gated idle protocol, EEGNet recorded 0.0% idle false activations. Why is that single number misleading on its own?

- **A.** Its command detection was only 1.7% — it almost always abstained
- **B.** The protocol contained no idle trials to trigger on
- **C.** Idle protocols have 40 classes so 0.0% is chance

Show published result

EEGNet: 1.7% command detection at 0.0% idle false activation · ShallowFBCSPNet: 43.3% detection at 1.7% false activation

A silent decoder never fires falsely, but never executes commands either — idle safety and command detection must always be read as a pair.

[Inspect idle & command protocol →](https://bci.report/protocols/idle/)

06 / 06 Montage reduction · 8 ch → 4 ch

### When posterior EEG coverage on 40-class SSVEP drops from 8 channels to 4 channels on the same 70 people, what happens to FBCCA and EEGNet?

- **A.** Both stay unchanged — 4 channels carry identical spatial information
- **B.** Both decline, and FBCCA stays separated well above EEGNet
- **C.** EEGNet overtakes FBCCA once channel count halves

Show published result

FBCCA 79.1% → 72.3% · EEGNet 55.8% → 44.1% · 4-ch paired difference +28.2 pp (98.75% interval +22.8 to +33.9 pp)

Cutting posterior electrodes in half lowers accuracy across methods, yet the training-free filter-bank structure still preserves strong frequency separation.

[Inspect SSVEP · 4 channels protocol →](https://bci.report/protocols/beta-4ch/)

05 · No single number

## Why there is no overall score

Different EEG paradigms obey different physics: frequency-locked visual responses (SSVEP), brief time-locked peaks (P300), slow spectral shifts (sleep staging), and idle false-activation trade-offs do not share one scale. Averaging them into a composite leaderboard number would erase which methods succeed on which signal.

Instead of collapsing protocols, inspect how methods cluster across tasks on the interactive 2D Method Map and Protocol Agreement view, or compare all frozen foundation encoders side by side across the eight core protocols.

- [Open the 2D Method Map & Protocol Agreement →](https://bci.report/explore/)
- [View the Foundation-Encoder Overview →](https://bci.report/methods/#foundation)

06 · Next steps

## Where to go next

- [Explore the 2D Method Map](https://bci.report/explore/): Visual comparison · Plot method profiles across protocols, inspect where protocols agree or diverge, and compare any two methods side by side.
- [Browse the 17 Questions](https://bci.report/topics/): Evidence by question · Read focused answers on dry vs. wet sensors, VR transfer, calibration budgets, model adaptation, pretraining controls, and dataset audits.
- [Check the Open Register](https://bci.report/datasets/register/): Full provenance · Search every EEG dataset and model tracked by BCI Report, see its step on the 5-step verification ladder, and inspect recorded terms.

[Looking for a definition? Browse all 36 terms in the bilingual Glossary →](https://bci.report/glossary/)

---
Markdown copy of https://bci.report/start/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/
