Removing the test correction did not give a 10 percent error gate in eight offline SEEG recordings under a binomial trial model
We compared four confidence checks in eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. The decoder, five thresholds, and later test data were fixed. All 80 joint checks at a 10 percent error target failed at test level 0.05. Removing the test correction therefore did not give a gate on this grid. At a 20 percent target, both fixed sequences retained 150 later windows with eight wrong labels. All outputs came from one person. The prior long-calibration gate retained 123 windows with three wrong labels. The prior short-calibration gate retained none. The package contains a proof, three figures, complete features, pinned raw-data retrieval recipes, code, and independent numerical checks. The formal guarantee requires independent trials and stable state distributions. The blocked task and reused cohort do not establish a live or prospective error limit.
stereoelectroencephalography · brain-computer interface · selective classification · risk control · fixed sequence testing · binomial test · calibration · offline evaluation
References
5- usesNo tested confidence threshold met both 10 percent error bounds in eight offline SEEG recordings sci:sha256:6298d8b9e5cb9e57a07b1da9336a42a3b12cf06bfa99e37c74670dfa693115c1
We used the exact decoder, feature archive, exclusions, split rules, and baseline gates for the matched comparison. All 32 prior gates, all model states, and all trial scores were reproduced.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
We used its complete public SEEG features, trial records, source metadata, and version-pinned raw byte ranges. We checked 118 inherited file hashes and independently checked stored and fresh source ranges.
- usesOffline SEEG label sets had 20.2 percent mean single-label error at a nominal 90 percent coverage target in eight recordings sci:sha256:e76d247dd4bb100a7cd2d2a24e80039a05928e303c3b41567c618d9770022089
The inherited decoder uses its training-only pair selection, scaling, and NumPy Newton logistic fit. This fixed model made a direct comparison of confidence tests possible.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used the exact version 1.0.2 SEEG dataset and its recorded trial data, features, source metadata, and version-pinned raw data ranges.
- doi:10.1088/1741-2552%2Facae0a
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesFixed early confidence thresholds reduced mean task state error to 24.5% at 49.6% retention after condition changes in eight offline SEEG tests
This strict error gate study defined a relevant limit. We tested mean error reduction and did not claim a certified error bound.