A fixed SEEG decoder gives higher observed mean accuracy with mixed samples than separate blocks for one film in eight adults
We tested speech and music prediction in eight adult SEEG records from OpenNeuro ds003688, version 1.0.7. A fixed decoder used five neighbors and high gamma power. Each record had 312 separate one-second windows. Mixed sample tests gave 57.91% mean balanced accuracy. Tests on separate film block pairs gave 52.84%. The observed difference was 5.07 percentage points. Its 95% subject bootstrap interval was 0.65 to 10.34 points. 7 patient effects were positive. The exact sign-flip value was 0.0703, under independent symmetric effects. Arbitrary block labels gave 56.01% in mixed tests and 49.91% in block tests. Recording time alone gave 99.99% and 16.67%. The package preserves source hashes, features, test predictions, split definitions, code, and checks. The tests measure different prediction tasks. Their difference does not isolate temporal dependence or a speech mechanism. Population inference assumes independent suitable patients. The result applies to these eight records and one film.
SEEG · speech decoding · music · balanced accuracy · cross-validation · temporal dependence · neural signals · epilepsy
References
2- usesdoi:10.18112/openneuro.ds003688.v1.0.7
This analysis used the SEEG recordings and metadata from OpenNeuro ds003688 version 1.0.7.
- doi:10.1038/s41597-022-01173-0
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesEarly selected SEEG pairs did not detect shared high gamma timing within later blocks of one film in eight adults under a circular shift null
We reused its eight adult SEEG feature arrays, cohort rule, source metadata, full raw provenance, and retrieval code. We verified inherited hashes and fresh raw windows. The new test selects contacts in early film blocks and measures shared timing in later blocks after each block mean is removed.
- usesPooled high gamma signals had a mean correlation of 0.016 in eight film SEEG records after block mean removal
We used its checked power features, cohort selection, source metadata, raw windows, and retrieval code. We tested shared timing in the same records after block mean removal. This is a new analysis of the same cohort.