No tested confidence threshold met both 10 percent error bounds in eight offline SEEG recordings
We tested five fixed confidence thresholds in eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. The data contain 2,014 retained trials. Each trial has a rest window and a window before a key response. We compared short and long calibration sets. The fitted decoder and 1,348 later test windows stayed fixed. Each gate needed a separate exact binomial upper error limit for each true state. Ten checks shared a 0.05 error budget per person and calibration length. No threshold met both 10 percent limits with either calibration length. At the secondary 20 percent target, one gate passed with long calibration. It issued 123 test outputs and made three errors. An observed-error rule passed at 10 percent in three recordings with long calibration. One sparse gate then made two errors in two accepted response windows. With zero errors, this bound procedure needs at least 51 accepted calibration examples per state at the 10 percent target. Task mix changed between calibration and test in all recordings. The binomial result requires independent trials and stable state distributions. These data do not establish a live command error bound. The package preserves features, source records, code, all candidate checks, three figures, and independent validation. This is a new analysis of a reused clinical sample.
selective classification · error bounds · calibration · stereoelectroencephalography · brain-computer interfaces
References
5- 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 original feature definitions, all 2,014 retained trial features, event records, exclusions, and version-pinned raw range recipes. These inputs allowed a matched calibration comparison and fresh raw checks.
- 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
We used its complete source archive, features, metadata, extraction code, and training-only decoder. We fit new first-third models. We tested separate accepted-error bounds with fixed short and long calibration sets.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed features and source records derive from OpenNeuro ds004473 version 1.0.2, whose recorded dataset metadata declares CC0.
- doi:10.1088/1741-2552%2Facae0a
- supportsdoi:10.1093/biomet%2F26.4.404
This is the Clopper–Pearson exact binomial method used for the state-specific upper error limits.
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesRemoving the test correction did not give a 10 percent error gate in eight offline SEEG recordings under a binomial trial model
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.