Leakage-controlled baselines on 3 public imagined/silent speech datasets before building a 16-ch rig
Med student building a 16-ch ADS1299 rig + 3D-printed headset from my own CT for imagined speech. Before recording I ran linear, leakage-controlled baselines on Gaddy (sEMG), Nieto and FEIS (EEG).
What came out:
- sEMG voiced → silent transfers (small but above the permutation null).
- EEG spoken → imagined does not, on Nieto or FEIS.
- Imagined speech collapses across days (Nieto: within-session 34%, leave-one-session-out at chance).
- Same-trial splits fake a transfer effect (FEIS 8.7%, p = 0.0002) that temporal-block splits remove (6.4%, chance). Same classifier, only difference is temporal proximity.
Repo with code, tables and feature caches (MIT, DOI on Zenodo):
https://github.com/gilbertoluporini-create/imagined-speech-baselines
Would love to hear from anyone with Ultracortex / dry-electrode experience on electrode repeatability across sessions. That is now the first gate of the project, before any decoding.
Comments
Correction to this thread: the three Nieto (ds003626) "sessions" were recorded consecutively on the SAME DAY, not on different days. So the collapse is across same-day sessions (cap presumably kept on): harder for the field, but it says nothing about day-to-day drift. README fixed.
The same-day correction matters. Would you try a remove-and-refit session alongside one where the cap stays on? That could help separate placement effects from changes between recordings.