EEG_MI — Motor imagery on a 32-channel dry-electrode cap

Mirror of github.com/twu3202/EEG_MI. Data lives in Twu31/cap32-mi-eeg.

Decoding motor imagery (MI) from a 32-channel dry-electrode EEG cap (TI ADS1299, 250 Hz, WiFi/UDP), with two goals: reach usable accuracy with minimal per-subject retraining, and find out whether EEG foundation models actually help on hardware like this.

Everything here is measured on one cap and one subject. The reports say plainly what the data supports and what it does not.

What we found

Question	Answer
Left hand vs right hand	Not decodable (p = 0.41). A dry cap does not resolve C3 vs C4 well enough.
Both hands vs rest	Decodable, AUC 0.83–0.87 (n = 19, breadth search over 115 pipelines)
Hands vs feet	AUC 0.704, p = 0.040 in mu (8–13 Hz) on 17 central+frontal channels — exploratory, selection-biased
Do foundation models beat classical?	No. Seven frozen backbones, none beat CSP/Riemannian tangent space + LR (0.790)
Channel-to-channel crosstalk	Not detectable; bounded at −24 dB by a measurement whose floor was set by open inputs

The single most useful negative result: the first foundation-model ranking was an artifact of zero padding. Feeding a 3 s trial to a checkpoint built for 15 s made 80 % of LaBraM's input zeros, and the probe read the padding. BENDR's "best-in-class" 0.733 became 0.303 once the input was fixed. src/foundation/embed_health.py is the pre-flight gate that now has to pass before any probe score is interpreted.

Reports

docs/mi_pilot_report.pdf	MI pilot: ERD/ERS, breadth search, foundation-model benchmark (中文)
docs/crosstalk_report.pdf · _zh	Crosstalk measured with an external generator (EN / 中文)
docs/hardware_acceptance.pdf	Noise, DC, mains, crosstalk acceptance suite (中文)
docs/impedance_injection_report.pdf	Reverse-engineering the 31.2 Hz impedance injection (中文)
docs/network_setup.md	Recording over the cap's WiFi AP without losing internet
research/	MI + foundation-model survey, per-model input contracts, OpenBCI notes

Data

The recordings are published separately, with a data card and a standalone loader: huggingface.co/datasets/Twu31/cap32-mi-eeg

Raw µV, unfiltered, pre-CAR, all 32 channels retained — including the ones that were dead in that session. Read the data card before analysing: one session lost 31 of its 50 trials to a receiver stall, and another has a visual confound during imagery.

Layout

src/
  common/montage.py          32-ch 10–20 montage + ADC scaling (µV = counts × 0.02235)
  acquisition/               cap → GUI/LSL: framing, MI paradigm, impedance, hardware tests
  experiment/mi_paradigm.py  full-screen cue window (fixation → cue → imagery → rest)
  analysis/                  epoching, artifact handling, breadth search over pipelines
  foundation/                frozen-backbone probing + the representation-health gate
  baselines/                 CSP / Riemannian baselines on MOABB (BCI IV-2a/2b)
docs/       reports (LaTeX + PDF); preamble.tex is shared by all of them
research/   surveys and notes written before the hardware arrived
results/    figures and metrics reproduced by the scripts above

Getting started

conda create -n eegmi python=3.11 && conda activate eegmi
pip install -r requirements-cpu.txt      # baselines, analysis (no GPU)
pip install -r requirements-dl.txt       # torch (MPS) + braindecode + foundation models

Without hardware, everything still runs on the synthetic source:

python src/acquisition/cap_gui.py --source synth

With the cap: join its ESPBCI access point, set a static IP of 192.168.4.2 (see docs/network_setup.md), then

python src/acquisition/cap_gui.py --source udp

Notes on the hardware

The cap is a low-cost DIY-class device, and several of its documented behaviours were wrong or absent, so they were measured rather than trusted: the frame layout, the lowercase command set, the 31.2 Hz impedance injection current (~24 nA, not µA), and the vendor's two mutually inconsistent impedance formulas. src/acquisition/ contains the test scripts for each. Vendor manuals and software are not redistributed here.

License

MIT for the code. The recordings are CC BY 4.0 — see the dataset card.