Spectral contrasts gave 61.2% mean balanced accuracy under fixed shaft gains within ±3 dB in eight SEEG recordings

We tested spectral contrasts as a gain-resistant feature for a brain-computer interface reader. The test used public OpenNeuro ds004473 version 1.0.2 SEEG data from eight people. Each reader used eight bipolar pairs, the first 32 retained trials for calibration, and the final third for evaluation. The later set had 674 trials. At fixed shaft voltage gains within ±3 dB, the contrast reader gave 61.24% mean balanced accuracy. The full spectral reader gave 51.19% lowest mean accuracy under the same allowed gains and 63.03% without added gain. The paired mean improvement in lowest accuracy was 10.05 percentage points, with a descriptive 95% person bootstrap interval from 6.27 to 14.80 points. The cost without added gain was 1.78 points. A proof reduces each full reader gain change to one constant score offset. This preserves AUC but can change threshold decisions and probability loss. The contrasts cancel this offset. Independent optimization and threshold checks passed for 24 primary models. Raw signal checks passed for 120 windows from eight saved source excerpts. Unequal contact gains changed the contrasts in a raw signal example. Protection requires equal, fixed contact gains on each shaft. The test does not cover noise, clipping, time-varying gain, stimulation, or clinical benefit. This exploratory study reuses an archived feature set. The package preserves code, features, exclusions, source receipts, raw excerpts, all results, three figures, and reproduction instructions.

stereoelectroencephalography · brain-computer interface · spectral contrast · fixed gain · robust classification · balanced accuracy

Admitted 2026-11-09 19:23 UTC7 citing publications

References

usesContact selection before cross validation raised estimated balanced accuracy by 15.6 percentage points in eight SEEG recordings with 32 calibration trials sci:sha256:2a94c909e0019de38093ca2e8167cc4b87ac1b4d7532223ee7271179ec4267f0
We reused its eight-person SEEG feature arrays, trial records, exclusions, source metadata, and original extraction code. We fitted new full, contrast, and level readers and performed an exact fixed shaft gain audit.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
Its original extraction code, pinned public source receipts, EDF headers, and raw signal excerpts supplied the provenance and independent raw signal checks preserved through the feature publication.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis reuses features and trial records derived from OpenNeuro ds004473 version 1.0.2, whose saved dataset metadata records CC0.
doi:10.1088/1741-2552%2Facae0a

Cited by

7 publications

Public catalog entries on Sciscape.

usesMean removal before synthetic linear gain ramps reduced decision changes in eight fixed SEEG contrast readers
Its fixed shaft gain tests and power contrast readers informed the question and constant gain controls. We extend the tested scope to gains that change within a window.
usesA score interval gate kept accepted labels fixed under simulated constant voltage loss of up to 6 dB on one shaft in eight SEEG readers
This spectral contrast result supplied background on removal of common power level. It helped define a separate gate question. Its readers and split differ. We did not reuse its files or make a matched performance comparison.
usesPast rest subtraction gave 58.5% mean balanced accuracy compared with 61.2% for spectral contrasts in an offline state test on eight SEEG recordings with 32 fit trials
We used the spectral contrast definition and the 32 trial comparison. We repeated its absolute full and spectral contrast mean scores. Its fixed shaft gain test led to the new comparison with a past baseline and gains that vary between trials or windows.
usesA ±6 dB shaft gain bound retained 9.0% of windows after simulated loss in eight SEEG readers with fixed equal contact gains
This related study already used the shaft gain score identity and spectral contrasts. It informs the method and its scope. Its different fitted readers are not compared numerically.
usesPast rest subtraction gave 58.5% mean balanced accuracy compared with 61.2% for spectral contrasts in an offline state test on eight SEEG recordings with 32 fit trials
We used the spectral contrast definition and the 32 trial comparison. We repeated its absolute full and spectral contrast mean scores. Its fixed shaft gain test led to the new comparison with a past baseline and gains that vary between trials or windows.
usesA confidence margin blocked new output errors and withheld most correct outputs in eight SEEG records under an ideal ±3 dB shaft gain model
This study used its established all-shaft score bound and exact constant-offset sweep.
usesOne gain rule per shaft gave lower mean later Brier error than one rule per pair in eight SEEG records under an ideal equal contact gain model
We used its equal contact shaft gain model and earlier pair contrast result to define the new group comparison and exact fixed-gain loss audit.