Mean subtraction before early bin removal gave 7.9% site accuracy with zero late SEEG inputs in five offline patient tests
Mean subtraction can move information from removed time bins into retained SEEG bins. We tested this effect in OpenNeuro ds004457 version 1.0.2, with 183 stimulation sites in five patients. Each site supplied four fitting pulses and four test pulses. We set the 60–100 ms input bins to zero before mean subtraction. Subtracting the 20–100 ms mean before early bin removal still gave 7.9% equal-patient mean site accuracy. Removing early bins before mean subtraction gave 2.8%, equal to uniform chance. With the measured late inputs, the two rules gave 28.2% and 19.6%. A pre-pulse baseline rule gave 24.4%. The full mean minus late mean gain was 8.55 percentage points; the five-patient two-sided sign-flip p value was 0.0625. An exact identity and an input perturbation prove the path from removed bins to retained bins. The site schedule limits claims about neural information or future BCI use. The package supplies three figures, source bins, all predictions, code, source hashes, and reproduction instructions. Fresh raw extraction and analysis replay matched all preserved arrays and results.
SEEG · intracranial EEG · mean subtraction · time-window leakage · stimulation site decoding · brain-computer interface · signal processing
References
6- usesPulse time alone gave 74.8% mean site accuracy in five offline SEEG tests with shared calibration and test blocks sci:sha256:07f3c9173d8e02cafb724475eb4d7d43bb6e6bf6e042867f73744b4ea8895b75
Its time-only site result and source block audit limit neural and online interpretations of this same offline pulse split.
- usesTraining with nine window shifts raised mean site accuracy under uniform −20 to 20 ms shifts but reduced accuracy at zero in five SEEG sessions sci:sha256:ddcdfc4be7aa6cee291db41c5f68acf1b6167edea4c15f5e92cea8e15dbac679
We reused its source-derived voltage bins, pinned raw file recipe, recording contact rules, and first-four versus last-four pulse split. We rebuilt the bins from raw files and added an interval mean audit.
- usesRetained covariance gave 29.1 percent mean stimulation site accuracy after known simulated contact loss in five SEEG sessions sci:sha256:7d9856cbd13da084e7110eb2e7e7c7fcab41e64cbdb5b077cfc162a230169490
We used its fixed Gaussian covariance specification with 0.5 diagonal shrinkage. We fitted all models again for the new interval input rules.
- usesSEEG response shapes at 15–100 ms gave higher site identification accuracy than shapes at 100–500 ms within each of five patient sessions sci:sha256:aa5b71d76d214f861e99964287f9da01f3f9fc6c707a4dcfe24807283918ab40
Its separate-window centering informed the scope of our question. Our test concerns mean subtraction before bin removal and gives no basis to change that earlier separate-window result.
- usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis used OpenNeuro ds004457 version 1.0.2; its preserved metadata identify this version and state CC0 terms.
- doi:10.1523/JNEUROSCI.1325-22.2023
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesAn 80 ms SEEG window at 100 ms gave lower site accuracy than at 20 ms in five offline 6 mA sessions
Its mean-subtraction leakage result informed our strict window preparation. We select available bins before centering and verify that unavailable bins cannot change features.
- usesPair selection at 200 to 500 ms did not show a clear SEEG AUC gain with five calibration pulses in eight sessions
Its centering-order result informed input isolation. This study removes samples before mean subtraction and checks that excluded samples cannot change scores.