AEA — Auto EEG Analysis

Drop your EEG data. Get a paper-ready figure, a COBIDAS methods section, and a Reviewer 2 memo — before you submit.

GitHub Repo Why AEA exists
AEA output

Auto-generated from raw .fif — MNE sample data, auditory vs visual N100

Verified

Auto methods section

Reads pipeline logs, writes a COBIDAS-MEEG paragraph. Every number from disk, not memory. No other EEG tool does this.

Verified

Reviewer 2 simulator

GPT-5.4 plays a hostile EEG reviewer. Found 4 critical + 9 major issues in our case study. We fixed them before any human saw the work.

Core

Recipe library

Each recipe = one landmark paper's analysis, made executable. N170, P300, alpha-attention, ERN-flankers — 4 recipes and growing. Like HuggingFace for EEG.

Three workflows

W1 · Recipe

Reproduce a published analysis

Pick a recipe, point at your data, get a figure + methods + citation + audit. One command.

/eeg-recipe n170-face-inversion --data projects/my-study
W2 · Pipeline

Custom analysis

Define your claims and contrasts. AEA auto-detects data shape, asks 3–4 questions, runs everything.

/eeg-pipeline projects/my-study
W3 · Audit

Pre-submission review

Cross-model: GPT-5.4 reads your plan + results + figures. Writes the rejection memo you'd get from Reviewer 2.

/eeg-audit projects/my-study --reviewer-mode adversarial

Skill map

19 Markdown skills (6306 lines, ~38 papers distilled). No pre-written code — the LLM reads each SKILL.md and generates MNE-Python code on the fly.

SkillWhat it doesStatus
eeg-preprocessFilter, re-reference, bad channelsVerified
eeg-icaICA + ICLabel auto-labelingVerified
eeg-epochSegmentation + artifact rejectionVerified
eeg-erpCondition averaging, difference wavesVerified
eeg-tfrTime-frequency (Morlet wavelets)Verified
eeg-statsCluster permutation (channel x time x freq)Verified
eeg-figurePaper-ready multi-panel figuresVerified
eeg-methods-textAuto COBIDAS-MEEG methods paragraphVerified
eeg-auditCross-model reviewer (GPT-5.4)Verified
eeg-spectralPSD, FOOOF/specparam, IAPFVerified
eeg-decodingMVPA, temporal generalizationVerified
eeg-behaviorRT/EEG correlation, single-trial regressionVerified
eeg-connectivitywPLI, PLV, coherencePlanned
eeg-sourceSource localization (dSPM, eLORETA)Planned

Recipe library

Each recipe encodes domain expertise from a landmark EEG paper into an executable Markdown file.

RecipeParadigmSource
n170-face-inversionN170 face vs objectRossion et al. 2003
p300-oddballP300 to rare targetsPolich 2007
alpha-attention-cueingPosterior alpha desyncWorden et al. 2000
ern-flankersERN + N2 + frontal thetaKappenman 2021 / Gehring 1993

Contribute a recipe or skill → significant contributors are invited as co-authors on the AEA paper.

Quick start

# Install git clone https://github.com/dengzhe-hou/auto-eeg-analysis cd auto-eeg-analysis conda env create -f environment.yml && conda activate aeais bash tools/env/check_env.sh # Use claude > "Analyze my N170 experiment in projects/my-study"

AEA auto-reads file headers, asks only what it can't infer, then runs everything.

Case study results

3 completed case studies validate the full pipeline.

Case study 1: MNE sample (auditory N100)

p = .0002
Cluster permutation
(5000 perms)
d = −0.46
Cohen's d
(medium effect)
90.3%
Epoch retention
(289/320)
FAIL → PASS
Audit loop
(9 issues fixed)

Case study 3: ERP CORE Full (3 claims, 6-panel figure, 94.6s)

ERP CORE Flankers (Kappenman 2021) — BioSemi 30ch, 1024Hz. 3 pre-registered claims (N2, ERN, frontal theta), Bonferroni corrected.

ERP CORE Full case study — 6-panel figure

Auto-generated 6-panel figure — ERN + N2 + frontal theta, from raw BDF in 94.6 seconds

3 claims
N2 (stim), ERN (resp),
theta (TFR)
94.6s
Raw to publication
figure
6 panels
Auto-composed
multi-panel figure
13 stages
Fully automatic
pipeline

How AEA relates to other tools

AEA = ARIS architecture + EEG domain expertise
↓ generates code for ↓
MNE-Python (computation)

AEA is to MNE what HuggingFace is to PyTorch