A 32-trial rest covariance update did not establish higher SEEG balanced accuracy than a mean update in eight people
We tested whether past rest covariance could improve a fixed SEEG task-state reader beyond a past rest mean update. We used public OpenNeuro ds004473 version 1.0.2, eight people, eight contacts per person, and 674 later trials. Both updates used 32 preceding rest periods. The covariance rule used 0.5 shrinkage toward the identity matrix. Full covariance updating gave 70.43 percent mean balanced accuracy. Mean updating gave 69.58 percent. The change was +0.85 percentage points. The descriptive 95 percent person bootstrap interval was [-0.20, +1.73] points. The exact two-sided sign-flip p value was 0.164. This evidence did not establish an accuracy gain. Full updating slightly reduced AUC and increased log loss. The package preserves code, feature arrays, source ranges, decisions, update maps, and three figures. Fresh source audits and independent numerical checks passed. This is an exploratory offline test of event windows. It does not test continuous BCI use or brain stimulation.
stereoelectroencephalography · brain-computer interface · rest update · covariance alignment · balanced accuracy · OpenNeuro · offline evaluation
References
5- usesJoint selection gave 68.8% later balanced accuracy and lower mean feature correlation at eight contacts in eight SEEG recordings sci:sha256:5f2909fee9833e95bc2b3075516752bcc5e65dcf73a3ecb2369673a2dcf79db4
Used its verified feature package, chronological split rule, and greedy additive panel code to construct matched frozen readers with eight contacts.
- usesA ten-trial rest update did not establish higher later-window SEEG accuracy in eight patients at thresholds set for 10 percent rest false positives sci:sha256:a2ab4f25dd9e0d622d6cfafdf4e6924fd0589a02a21912150ae1a6c813a707b0
Used its scalar rest-update result to define the covariance-versus-mean comparison. Its accuracy values are not treated as matched estimates.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
Used its ds004473 version 1.0.2 log-power arrays, event definitions, eligibility rules, pinned source ranges, and extraction recipe.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source feature arrays, task metadata, and retained trial records came from OpenNeuro ds004473 version 1.0.2.
- doi:10.1088/1741-2552%2Facae0a
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesEarly fits did not establish lower loss than full rest correction with a 10-trial half-life in eight SEEG readers
Its comparison of rest mean and covariance updates did not establish a mean later accuracy gain. It motivated a bounded-strength check. It uses another contact panel and split, so it is not an independent replication of this result.