Training for ±6 dB gain on one shaft raised correct classes with a guarantee in eight SEEG records under a linear gain model
We tested whether training for an unknown shaft gain improves a small SEEG task decoder. We used OpenNeuro ds004473 version 1.0.2 and eight clinical records. Models learned from early trials and classified rest and pre-response windows in 674 later trials. The allowed change was a common voltage gain from −6 to +6 dB on at most one shaft per window. It assumed linear acquisition, no clipping, no added noise, and unchanged labels. An exact score interval gave a proof that an accepted class stays fixed. Gain training raised the equal-person correct fraction with this proof from 9.6% to 68.7%. The paired gain was 59.2 percentage points. Its descriptive 95% bootstrap interval was 47.7 to 70.8 points. All eight differences were positive. The exact two-sided sign-flip p value was 0.0078125 under the stated null assumptions. The gain fit accepted all windows but was wrong on 31.3%. Its intact accuracy was 68.7%, compared with 69.7% for the standard fit. Independent solvers and complete fresh source reconstruction passed. The cohort was reused, so this is exploratory evidence. The result proves class stability under the signal model. It does not prove live control, fault detection, class correctness, or a stimulation effect. The package includes features, pinned source recipes, fitted models, scores, figures, code, and a scoring tool that needs no labels.
stereoelectroencephalography · brain-computer interface · robust classification · gain invariance · logistic regression · signal stability · convex optimization
References
5- usesA distance monitor missed 54.4% of new errors on average under simulated ±6 dB shaft gains in eight SEEG recordings sci:sha256:3901ecb49a9e150042b75bc7d9429399128e7986d37dd5fdeae1159a3ab831c2
We used this package for the eight-person cohort, exact power features, chronological split, pair rule, source range recipes, extraction code, and standard logistic solver. We checked all parent files and rebuilt all source features. We compare new gain training with its standard fit.
- usesBand power ratios retain 67.0% balanced accuracy under simulated constant voltage gains per shaft in eight SEEG recordings sci:sha256:fa4b2be456049aafc97572ea0a4d44239abb9b5dc00341d8bcb7a0fd368bf201
This earlier study tested power ratios on the same cohort. It establishes the earlier use of invariant power features. We use its report to state the scope of the new training experiment. We make no matched performance comparison with its fits.
- usesPower contrasts preserve fixed decisions under simulated broadband gain changes in eight SEEG recordings sci:sha256:68da70f80a5598956f61bc64b69eb4a62fc811e861e263de58c286a9e941c775
This earlier study proved gain invariance for power contrasts. We use that result to explain the invariant subspace reached by the bounded gain loss. We do not claim a new power-contrast theory.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used OpenNeuro ds004473 version 1.0.2 as its source dataset.
- doi:10.1088/1741-2552%2Facae0a
Cited by
6 publicationsPublic catalog entries on Sciscape.
- usesEndpoint tests missed seven decision changes between simulated contact gains 0.8 and 1.2 in eight SEEG records
We use its shaft gain bound as prior proof scope. This study gives a separate bound for nonlinear band power changes at one recorded contact.
- usesQuadratic bounds certified 39.4 percent of later SEEG decisions for simultaneous constant contact gain changes of up to five percent in eight recordings
Its common shaft gain certificates informed the distinction between additive log-power uncertainty and unequal contact gains before bipolar subtraction.
- 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 score interval gate for class stability. The source split differs from this study.
- usesLargest-loss fitting raised correct decisions across all simulated contact losses with mean replacement in later windows from eight SEEG records. Intact accuracy fell.
Its earlier worst-case shaft-gain training test informed the novelty assessment. Our test uses physical contact loss, mean replacement, and a different split and objective.
- usesLinked band bounds proved more SEEG classes stable under 5% constant digital contact gain errors in eight records
We used its stated common-shaft linear gain assumptions to define the scope difference. The present contact model allows unequal gains before bipolar subtraction and is not the earlier common-shaft model.
- usesEndpoint tests missed the largest accepted error in 19 of 35 shaft cases for eight fixed SEEG readers under ideal constant gains within ±6 dB
We use its affine shaft gain interval and class stability work as method context. We add exact confidence-gate error ratios and label-aware separate-window bounds. No model or numerical comparison is reused from that study.