Targeted feature changes on one shaft caused wrong outputs from all 507 withheld windows in eight fixed SEEG output gates
We tested a shaft omission gate in eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. The gate gave a label only when all fixed readers agreed. Each extra reader omitted one selected electrode shaft. A proof showed that one changed shaft could not reverse a clean accepted label. The gate retained a mean 67.3% of the full readers' correct calls on later trials. Its clean conditional error was 23.0%. But targeted feature changes caused wrong accepted labels from all 507 withheld windows. Each change affected one shaft and used the true state label. Separate windows used separate changes. No clean accepted correct label could become a wrong output under the proof assumptions. Ideal gain changes also gave new wrong outputs from clean withheld windows. These feature tests do not measure physical fault prevalence. Fresh source ranges reproduced 32 trials exactly. An independent spectral check and a repeated analysis passed. The package preserves the proof, gate code, inputs, models, attack witnesses, source records, results, figures, and repeat instructions. This is an exploratory offline test in reused clinical data. It does not test live commands or brain stimulation.
SEEG · brain-computer interface · output gate · electrode shaft · feature perturbation · logistic regression · fault tolerance · offline benchmark
References
5- 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 sci:sha256:20ce0f1e203dd2355368851656294c8b747a2797464d2945044429e0c6d6c8ba
Its bounded voltage loss test motivated our question about a gate that needs no bound on a feature change. Its split and guarantee differ from ours. We did not compare their numerical rates.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This work supplies the feature files, trial rules, time split, and baseline reader. We used these inputs and reproduced the full reader before we tested the shaft omission gate.
- usesRank bounds certified 79.7% of equally weighted state decisions in 28 SEEG patients against any finite change to one nonnegative shaft score with fixed calibration sci:sha256:207a598f7dfa5076055d19c7ffb97570627d8433f6ceb07c38ec70501a6e58a9
Its median rank certificate provides context for shaft fault guarantees. It has a different decision rule and fixed calibration assumptions. Our result concerns the unanimity gate only.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed feature files and metadata derive from OpenNeuro ds004473 version 1.0.2; the publication preserves the source records and retrieval steps.
- doi:10.1088/1741-2552%2Facae0a
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesSix votes from eight shaft readers gave 5.04 percentage points more mean correct output than agreement among three groups in eight offline SEEG records
Its one-shaft feature test showed that agreement among readers that omit one shaft can admit wrong labels on clean withheld inputs. That result motivated separate input groups and a guarantee that covers every input. The prior gate is not a matched empirical control here.
- usesFive of seven local SEEG votes kept the clean majority under one shaft fault in eight recordings
Its one-shaft feature tests showed that an agreement gate could give wrong output from clean withheld windows. We used that result to define a protected clean-majority reference and to test withheld windows in the new local vote gate.