Causal feature re-centering did not improve later-trial SEEG key-press detection AUC in 8 subjects
We tested label-free causal re-centering of high-gamma (70–150 Hz) and beta (13–30 Hz) log-power features in offline cued SEEG key-press detection. For each trial, re-centering subtracted the mean of rest and movement windows from the previous K trials. A shrinkage LDA trained on trials 1–100 was evaluated on trials 101+ in eight subjects using archived features from OpenNeuro ds004473 v1.0.2. The pre-registered primary comparison found no AUC change for K = 40: median AUC was 0.874 with re-centering and 0.869 with fixed features; mean change was +0.0008, 3/8 subjects improved, and the exact two-sided sign-flip p-value was 0.91. Other windows, running z-scoring, high-gamma-only features, and five-fold CV also showed changes near zero. In a post-hoc positive control, re-centering removed a synthetic common linear offset ramp: with a 2-SD ramp, fixed median AUC fell to 0.857 while re-centered AUC remained 0.874 (7/8 subjects improved, p = 0.016). In these single sessions, the recordings did not show a slow drift that reduced fixed-decoder AUC. The study was offline and limited to one session per subject, at most about 200 trials after calibration; multi-day drift and changing class ratios were not tested.
stereo-EEG · brain-computer interface · feature re-centering · label-free adaptation · AUC · shrinkage LDA · signal drift
References
4- Beta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
[Legacy relation: reproduces] The fixed-decoder transfer AUCs for HG+beta (trials 1-100 to 101+) are the same as the published values for all 8 subjects.
- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
Archived HG and beta features of 8 subjects and the numpy shrinkage LDA code; fixed-decoder AUCs reproduce its values.
- supportsIn 8 SEEG subjects, unsupervised threshold adaptation did not improve balanced accuracy over a fixed calibration threshold sci:sha256:52aaaf95f6585fe3fecccbdc3f4209da3752d1a2cd19e2b12680307adabe70ac
Feature-level label-free adaptation also gave no gain on trials 101+, which agrees with the null result for unsupervised threshold adaptation.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The archived feature arrays analyzed here were derived from this OpenNeuro dataset release.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- supportsPast-only normalization kept cross-patient SEEG recall decoding above chance in 18 patients
Label-free causal normalization from past trials again did not change AUC, here for a cross-patient memory decoder (median diff -0.0017, 95% CI -0.0044 to +0.0023).
- supportsNo detectable time-on-task change in the SEEG movement response or fixed-decoder AUC in eight subjects
No baseline drift was found, which agrees with the result that label-free re-centering did not change AUC.
- supportsRe-centering preserved SEEG detector accuracy under simulated gain drift, but its advantage over shape features was not significant
Causal re-centering on the previous trials had a small no-drift cost here too (BA -0.003, AUC -0.0025 for K = 20).