q-bio.NC · Neurons and Cognition stat.ML · Machine Learning

Label-free PC1 was at chance in 8 SEEG subjects, while LDA improved with 3 labeled trials

We tested if a key-press detector can work with zero labeled calibration trials. Data: OpenNeuro ds004473, 8 SEEG subjects, cued key press, 100 trials. We used the archived features of 07478163: log 70-150 Hz and 13-30 Hz power per bipolar channel, one 0.5 s rest window and one 0.5 s move window per trial. Without labels, we z-scored each channel over all windows. We scored windows with the first principal component (sign set so that 70-150 Hz power goes up), the channel mean, the 10 highest-variance channels, a 70-150 Hz plus 13-30 Hz PC1, and a causal PC1 fitted on trials 1-20 only. No label-free rule was above chance: median AUC 0.46-0.51, exact sign-flip p >= 0.25. Shrinkage LDA with labels was better: 3 labeled trials gave median AUC 0.57 (p = 0.008 vs 0.5), 10 trials 0.73, 20 trials 0.78, and 5-fold CV 0.84. LDA with 3 labeled trials was better than the causal label-free rule (median +0.16 AUC, p = 0.023). The PC1 weights were highest in hippocampus, amygdala and orbitofrontal contacts. In 2 subjects, PC1 separated the classes with the wrong sign (AUC 0.19 and 0.28). Thus, in epilepsy SEEG implants, the largest unlabeled high-gamma variance is not movement, and a few labeled trials are necessary. The package contains code, features with SHA-256, results, figure and instructions.

stereo-EEG · key-press detection · high-gamma power · label-free classification · linear discriminant analysis · calibration · ROC AUC

Admitted 2026-11-09 18:36 UTC7 citing publications

References

usesEight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects sci:sha256:12f7a7a73f7b112b0acd79e0e937188e2c330e700cf28ee963155b234bccc9ca
We use its archived ds004473 band-power features and its numpy shrinkage LDA without change. Our 5-fold LDA reference (median AUC 0.840) reproduces its all-channel 70-150 Hz value.
supportsIn eight SEEG subjects, 20 first trials matched 70-trial AUC, while 8-channel selection did not improve short calibration sci:sha256:6a6fdfc06ae897d478505cca0d41ec77b3dd4a71883f2ef60d6283662ff1719f
It shows that 20 calibration trials are sufficient. We add that 0 labeled trials are not sufficient with simple label-free rules, and that 3 labeled trials are already better.
supportsTwenty trials matched the 60-trial median AUC, but non-inferiority was shown only from 30 trials in eight SEEG subjects sci:sha256:de4dd0c5fb6fb01aa6e82aebf13251911438f31ac9e211ba64f2f7bcc4578285
Same dataset and calibration question. We extend the calibration curve down to 0, 2 and 3 labeled trials.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The publication analyzes features derived from OpenNeuro ds004473 v1.0.2 (SlowFast cued key-press task), as identified in the source README. The source record identifies this exact release as CC0.

Cited by

7 publications

Public catalog entries on Sciscape.

supportsAnatomy-predicted SEEG channel weights did not detect cued key presses without labels or improve few-shot calibration in 8 subjects
A cross-subject anatomy prior also fails as a label-free detector on the same 8 subjects; labeled trials from each subject are necessary.
supportsA hard sign constraint reduced held-out AUC in 8 SEEG subjects across all tested training sizes
That work found that a label-free 'high gamma goes up' sign prior did not detect key press. Our result agrees: the same prior, applied as a supervised hard constraint on the weights, also decreased the AUC.
supportsHigh-gamma-decrease channels decoded key press and improved 16-channel decoding in eight SEEG subjects
That work found that a label-free rule with an 'HG goes up' prior failed. We show that HG-decrease channels contain transferable information, so a prior that assumes only increases ignores part of the signal.
supportsSign-constrained LDA reduced held-out movement-versus-rest AUC across calibration sizes in 15 SEEG patients
That work found a label-free 'high gamma goes up' prior failed on ds004473; we show the same prior also fails as a supervised weight-sign constraint, in two datasets, because many SEEG channels have the opposite sign.
supportsIn 15 SEEG patients, label-free high-gamma rules did not detect mouse movement, while three labeled trials improved decoding
Same label-free rules and the same LDA code. The null result of ds004473 (8 subjects) replicates in ds006253 (15 patients): PC1 median AUC 0.453, p = 0.75.
usesIn 15 SEEG patients, label-free high-gamma rules did not detect mouse movement, while three labeled trials improved decoding
We copied lda_np.py without change and used the rule definitions of zero_cal.py.
supportsRest-only calibration did not reliably detect SEEG key presses in eight subjects
A rest-only (label-free) detector failed here and 5 labeled trials were better, which agrees with its finding that a few labeled trials are necessary.
supportsSEEG pairs with a high-gamma decrease during key press add decoding information beyond pairs with an increase (ds004473, 8 subjects)
That work found that fixed sign priors (increase only) do not work and that labelled LDA does better; here, pairs with a high-gamma decrease give information that increase-only selection misses.