Synapse: 16-channel cEEGrid ear-EEG for auditing EEG foundation-model transfer

Released with the anonymous submission Does It Really Transfer? Auditing Frozen EEG Foundation Models on Small Ear-EEG Data (under double-blind review).

23 adults (16 with a sound-sensitivity disorder, EXP; 7 controls, CTRL), 16 around-the-ear channels at 125 Hz, 1,874 trials across three passive auditory tasks (let 575, hlt 575, ast 610) and one visual task (pmt 114, not analyzed in the paper).

Files

File	Content
synapse_125hz_pre1s_post7s.pkl	Dict with records (92 subject-task records: subject, group, task, X of shape (n_trials, 16, 1000) float32 spanning [-1,+7] s, channel_mask, n_trials), meta, subject_quality. X is in volts (MNE convention, median per-trial channel SD 9.2e-6 V over [-1,+3] s). Raw after 1-50 Hz bandpass and 60 Hz notch, no normalization. Needs numpy >= 2.0 to unpickle.
synapse_125hz_pre5s_post7s.pkl	The same 92 records from -5 s, in volts (1501 samples to +7 s, 1876 samples to +10 s for the ast task), for the window sweep, positive control and positive example. Its [-1,+7] s slice equals the main pkl exactly. Its meta (n_samples 1501, duration_sec 12) describes the non-ast records only. The preprocessing_provenance field in both pkls calls the upstream -5..+7 s epochs "not released", but this file is that release.
demographics.csv	Subject ID, group, age in 5-year bins and sex for the 23 subjects.
results_96cell_interp.zip	Subject-level LOSO predictions (per seed) for all 96 grid cells and per-seed AUROCs for the 3 supervised baselines, failed channels interpolated (paper). Extract into src/benchmark/.
results_96cell_keep.zip	The same grid with failed channels kept in the input.
code.zip	Reproduction code (MIT license) and the precomputed outputs of the audit analyses (window sweep, channel controls, selection-aware tests, embedding control, positive control, positive example, power simulation). See src/benchmark/README.md.
synapse.json	Croissant 1.0 metadata with Responsible-AI fields.

Quick start

from huggingface_hub import hf_hub_download
import pickle
path = hf_hub_download('synapse-ear-eeg/synapse', 'synapse_125hz_pre1s_post7s.pkl', repo_type='dataset')
data = pickle.load(open(path, 'rb'))
rec = data['records'][0]
print(rec['subject'], rec['group'], rec['task'], rec['X'].shape)

To regenerate the tables of the paper from the released predictions (one conda environment, environment.yml in code.zip, see src/benchmark/README.md Part 4 for the figures and the full list):

unzip code.zip && cd src/benchmark && unzip ../../results_96cell_interp.zip -d .
mkdir -p ../../data && cp /path/to/synapse_125hz_pre1s_post7s.pkl /path/to/synapse_125hz_pre5s_post7s.pkl /path/to/demographics.csv ../../data/
python regen_tables.py && python regen_robustness.py   # Tables 2 (Interp.), 6, 7, 8, 14
python tables_selection.py                             # set A rows of Table 3, Table 15
python tables_audit.py                                 # Table 3 in full, Tables 4, 12, 18
python tables_positive_control.py                      # Tables 16, 17
rm -rf results && unzip ../../results_96cell_keep.zip -d .
SYNAPSE_BAD_CHANNELS=keep python regen_tables.py       # Kept columns of Tables 2 and 10

Failed channels

Channels that failed quality control are kept in the released signals and flagged in channel_mask. The mask is correlated with the label (the 16-d mask alone gives LOSO AUROC 0.76). The released loader (data.load_pkl) replaces each failed channel by the mean of the passing channels of the same ear by default, but the interpolated channels duplicate the ear mean, so the failure pattern stays recoverable. Compare models on the three channels that pass QC in every subject (L01, L08, L09) or drop failed channels.

Analyses added to the audit

src/benchmark/run_revision_analyses.sh lists the commands, with precomputed outputs included: pretrained-embedding control (Zp, Zt), selection over grid cells with all four probes, the positive control (recorded and injected evoked responses), the subject-level positive example (a pseudo-label with an injected transfer of known size, run through the full audit), and the power simulation. Rerunning them needs feature_cache/ from run_grid.sh and synapse_125hz_pre5s_post7s.pkl in data/.

Intended use and limitations

Intended as a testbed for methodology research on EEG foundation-model transfer. It is not a clinical screening dataset. The cohort is small and imbalanced in age and sex (EXP mean 41.7 y, 11M/5F; CTRL mean 25.3 y, all female), which limits absolute performance numbers. Used as scores, sex alone gives AUROC 0.84, exact age 0.74 (0.71 from the released 5-year bins) and the rule "male or older than 35" 0.91, so report these demographic baselines next to any model. The grand-average evoked response is weak: telling post-onset from pre-onset windows across subjects is at chance for every audited model and for band power. Questionnaire scores are not included.

License

Data, derived features and results: CC-BY-4.0 (LICENSE_DATA_CC-BY-4.0.txt). Code: MIT (LICENSE in code.zip).