Offline SEEG label sets had 20.2 percent mean single-label error at a nominal 90 percent coverage target in eight recordings

We tested public SEEG label sets in eight people from OpenNeuro ds004473 version 1.0.2. Early trials trained eight-pair decoders. Separate early trials set class-specific thresholds. Later trials supplied 1,348 test windows. At a nominal 90 percent coverage target, equal-patient mean coverage was 91.4 percent. Its descriptive 95 percent patient bootstrap interval was 89.9 to 93.2 percent. The sets gave one label in 49.6 percent of windows. Mean error among these single labels was 20.2 percent, with an interval of 13.5 to 27.3 percent. The pooled error was 115 of 678 single labels, or 17.0 percent. Reused training scores gave 83.7 percent coverage. A separate two-rest-position model gave 90.6 percent coverage but few useful decisions. These results show why coverage and selected-decision error need separate measures. The known rank guarantee requires exchangeability, which this test did not establish. The package preserves four figures, exact features, models, scores, code, source hashes, and retrieval instructions. Sixteen fresh raw ranges and independent numerical checks passed. The evidence concerns offline task windows within recordings. It does not establish live command reliability or stimulation effects.

stereoelectroencephalography · conformal prediction · label coverage · selective prediction · brain-computer interface · calibration · single-label error

Admitted 2026-11-09 18:47 UTC3 citing publications

References

supersedesOffline SEEG label sets had 20.2 percent mean single-label error at a nominal 90 percent coverage target in eight recordings sci:sha256:9e577d50245af75c7daaaabc8d7241929e4ec9341964b845fd606bcc1e139e47
This package corrects the file verifier to exclude the metadata folder that AutoSci adds during retrieval. The earlier package has the same features, models, numerical results, and figures. All ten numerical result files are unchanged.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This publication supplied the exact SEEG power features, trial exclusions, source metadata, and pinned raw retrieval records. We verified the imported hashes and fitted all new decoders from their training trials. We preserve the source code and the raw retrieval recipe.
usesEarly confidence scaling decreased mean Brier error from 0.193 to 0.185 in eight offline SEEG patient tests sci:sha256:f3d79f04937a343945e7a53f5b54b00a0284174fca051959a818f1c79d41a7d1
This result supplied the confidence-score context and temporal split used in this test. The present study tests separate class label sets and error among single labels. Both studies use the same public recordings, so they are not independent replications.
usesRest calibration gave a 4.2 percent mean false positive rate in eight later-trial SEEG tests sci:sha256:7a28b8fc519f8f608fe5546ba40889db9bb37b5c93f68ed3821b4ba5bf1c5e75
This result supplied a prior rest-threshold benchmark on the same features. It used a different temporal split and a 5 percent rest target. The present study tests sets for both labels, total and class coverage, output fraction, and error among single-label decisions.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
This is the exact OpenNeuro dataset version used for the source measurements, with its version-specific DOI and CC0 license recorded in source metadata.
doi:10.1088/1741-2552%2Facae0a

Cited by

3 publications

Public 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
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.
usesNo tested confidence threshold met both 10 percent error bounds in eight offline SEEG recordings
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.
usesA gate with five thresholds gave outputs in one of eight offline SEEG records at a 20 percent error target
Its distinction between label coverage and error among single-label outputs motivated the direct accepted-error target. We used its fixed fit/calibration/test groups and reader settings for a matched reference. We reproduced the 115 wrong single-label outputs out of 678 on the reused cohort.