Gain-invariant spectral features did not improve cross-session SEEG accuracy, while session centring did
In archived SEEG features from OpenNeuro ds004752, a shrinkage LDA decoder was trained and tested across 278 ordered session pairs from 10 subjects (56 sessions) to classify set size 8 versus 4. Spectral-shape features, which subtract each channel’s mean across six bands, did not improve cross-session balanced accuracy over raw log power (+0.003; 95% CI −0.010 to +0.021; p = 0.79). Label-free subtraction of each session’s mean feature vector increased accuracy by +0.027 (95% CI +0.011 to +0.044; p = 0.014; 8/10 subjects). Cross-session AUC (0.677) matched within-session AUC (0.675), consistent with a shifted fixed decision threshold. Between-session feature change was 2.2 dB RMS, with a gain-like fraction of 0.66 (chance 0.17; split-half floor 0.53; p = 0.004). Shape features reduced but did not remove the decoder score shift. This analysis covers one dataset, task, time window, six bands, bipolar pairs, and linear decoder; it includes 10 subjects, and centring used all unlabeled test-session trials.
SEEG · cross-session decoding · session drift · spectral features · balanced accuracy · shrinkage LDA · label-free centring · working memory
References
4- challengesIn 15 SEEG patients, simulated gain drift lowered load-decoding accuracy, while spectral-shape features preserved balanced accuracy sci:sha256:b6263c39028f94a46261aa5c8e9fe09fbb227be48cf4fbdab0053e242c0234aa
This work tests the simulated-gain-drift claim on real between-session drift: shape features gave no gain in cross-session balanced accuracy, but label-free centring did.
- usesCross-session SEEG decoder AUC declined slightly over 0–72 hours, with no clear day-boundary step sci:sha256:cc725dbcd3c0387d439ed6f12692fcc950e0c99a4d0ab3551ef59f010807b467
Input features (late-maintenance band power for each session) and session start times copied from this package, with the same SHA-256.
- usesPooling earlier sessions did not improve new-session SEEG load decoding over the latest session sci:sha256:6bb80018bacd7f34ffb73a0bbb89e2eb44676ae924aff2e4248e4c33a9b8d05e
Earlier archive of the same features and the shrinkage LDA code.
- doi:10.7554/eLife.78677
Cited by
1 publicationPublic catalog entries on Sciscape.
- supportsRe-centering preserved SEEG detector accuracy under simulated gain drift, but its advantage over shape features was not significant
Same direction in a different task: centring of full band-power features beat gain-invariant shape features (+0.041 BA, 7/8), but not significant after Holm correction here.