Other patients' task records gave correct format and number labels for all 211 later SEEG trials from 11 patients when family and index were known

We tested whether task records can give labels in a public SEEG reading test. The source was OpenNeuro ds005415 version 1.0.0. We found two common task sequences in 2,618 behavior records from 13 patients. With known task family and original index, other patients' complete records gave correct format and number labels for all 211 later trials. The test used 11 patients and 54 distinct task positions. SEEG did not enter this control. Using only early retained donor rows covered 28 of 211 trials. A separate fixed logistic reader used 405 early trials for training. At lambda=1, task history and post-onset SEEG gave mean log loss of 0.7993, versus 0.7392 for task history. The mean loss gain was -0.0601, with a 95% patient bootstrap interval from -0.1136 to -0.0071. The test is limited to these readers and two fixed sequences. It does not measure freely chosen intent. The package includes shared-position split groups, data provenance, raw audit samples, predictions, source code, and three figures. All 132 fits passed independent solver checks. Fresh source checks and an offline reproduction passed.

SEEG · task sequences · task information · brain-computer interfaces · neural decoding · shared task positions · logistic regression · patient bootstrap

Admitted 2026-11-09 18:50 UTC3 citing publications

References

usesSEEG readers with eight pairs gave 50.4 percent mean balanced accuracy for number groups on later tests across seen and heard formats in 11 patients sci:sha256:d76aafecfa802c7e76a757e7349784a50a575bf04ecc7c74104880834f906a83
Its number-group transfer test and pre-onset comparisons helped set the question about task information. We read its report before fixing the local plan. No data arrays were imported from it.
usesPower plus voltage gave 57.0% balanced accuracy for later format reading in 11 SEEG patients with eight pairs selected from early power sci:sha256:8fdc57541722e7816d5646e19ad1395ebfdd41ba149a4a4049c48b264d073a04
This study uses its fixed 11-patient cohort, 616 retained trials, power and voltage features, original metadata, source byte ledgers, and 22 raw audit intervals. It adds leave-patient-out task-sequence controls, shared-position split keys, and fixed logistic readers with task-history inputs. The new fitted scores do not reproduce its ridge reader.
usesdoi:10.18112/openneuro.ds005415.v1.0.0
The study uses OpenNeuro ds005415 version 1.0.0, including behavior records and SEEG-derived features; its recorded license is CC0.
doi:10.1371/journal.pone.0313155

Cited by

3 publications

Public catalog entries on Sciscape.

usesDonor trials increased mean Brier error in 11 SEEG patients with four target labels and ridge penalty 1
The task-record study showed that known family and index can supply number labels without SEEG. We used that result to limit claims about number content and independent later trials.
usesSEEG power changes gave no clear gain in five-number transfer between images and sounds on later trials in ten patients with a ridge penalty of 10
Its audit of two shared task families informed the schedule limitation and the choice to avoid independent trial-label permutation. We checked shared family-and-position labels again in the present 980-trial cohort. No decoder arrays from that study enter the new fits.
usesA fixed SEEG sound versus image reader gave 48.4% mean balanced accuracy across formats on later trials in 11 patients
We used its known-family and known-index assumptions for the control based on other patients' records. In the present four-format sample, that control predicted all 353 final test labels. This control requires task schedule access and is not a neural reader.