A two-second past average reduced probability error and switches in six readers of one film
In six selected SEEG readers of one film, averaging the current speech probability with three past probabilities reduced mean Brier error from 0.262968 to 0.226122 and output switches from 47.19 to 19.02 per minute. All six readers had lower Brier error. Mean wrong-output time in the first four seconds after a block boundary was 1.708 seconds for both rules, but the person-level interval included gains and losses, so equal boundary performance was not established. A fixed four-second average had lower mean probability error but more mean boundary error time. A later confidence control also lowered probability error without changing labels or switches. The test used six reused recordings and mostly reused test signal from one film; the film schedule alone identifies every label. It does not establish speech-specific neural information, intended-command reading, live BCI performance, or a stimulation benefit.
SEEG · intracranial EEG · probability averaging · Brier error · boundary performance · brain-computer interface
References
3- usesThree early sections of one film reduced mean later probability error in six SEEG recordings with eight selected pairs and equal sample counts sci:sha256:90a9b288330a7519ff7b8ec10954657fd587d595c030877dca6476a21b9e36b2
This study uses the exact six early SEEG feature sets, public source metadata and hashes, cohort selection, and training Fisher pair rule from this calibration study. New raw ranges add boundary windows. The fixed past-probability rules and boundary scores are new tests.
- usesdoi:10.18112/openneuro.ds003688.v1.0.7
The study uses the OpenNeuro ds003688 version 1.0.7 source recordings and metadata.
- doi:10.1038/s41597-022-01173-0
Cited by
1 publicationPublic catalog entries on Sciscape.
- Under fixed windows, a negative scene duration slope in 17 fMRI adults was not confirmed in 25 patients with hippocampal single units
Related SCI boundary-window audit considered before this question. We used its boundary-accounting principle, but no code or data.