Eight SEEG pairs showed no clear gain over fixation for letter load across sessions in 15 patients under assumed epoch timing

We tested whether a small SEEG readout predicts letter load in a later session. We used the first and last numbered session from each of 15 patients in OpenNeuro ds004752 version 1.0.1. We assumed that the public epoch starts at fixation and that session numbers give session order. Training used the first session. Testing used the last session. The model used eight adjacent depth-contact pairs and fixed power features. It predicted 4 letters versus 6 or 8 letters. The main test used 683 training trials and 703 test trials with the source clean flag. Mean patient balanced accuracy was 54.3% for delay change and 51.3% for fixation. The gain was 3.0 percentage points. Its 95% patient bootstrap interval was -0.9 to 6.9 points. The two-sided patient sign-flip p value was 0.165. Thus, this test showed no clear gain over fixation. It did not show that the true gain is zero. A model with trial history alone reached 55.2%. Delay accuracy was 52.4% when source flagged trials were included. The task made low-load trials more likely after errors. A session label shuffle does not preserve this rule and gives no valid task randomization test. The package has source records, retrieval code, features, models, predictions, figures, and checks. A full analysis rerun gave identical result files. One complete source hash was checked. The other 29 sources have S3 version and range-hash records. These findings apply to this offline task test. They do not establish a useful live BCI signal or a benefit from stimulation.

SEEG · verbal working memory · letter load · balanced accuracy · brain-computer interface · cross-session decoding

Admitted 2026-11-09 08:42 UTC9 citing publications

References

usesdoi:10.18112/openneuro.ds004752.v1.0.1
The analysis used OpenNeuro ds004752 version 1.0.1; its saved version-specific metadata declares CC0.
supportsdoi:10.7554/eLife.78677
The source task paper supplies the task timing and error-to-low-load rule used to define and assess the analysis windows and history checks.
doi:10.1126/sciadv.aav3687
doi:10.1038/s41597-020-0364-3

Cited by

9 publications

Public catalog entries on Sciscape.

usesAdded scalp power gave no clear mean depth beta prediction gain across sessions in 15 epilepsy patients after ridge calibration
The depth EDF parser and power extraction functions originated in this archive. We preserved its code notice and used its functions for fresh source checks.
usesA 12-trial confidence update lowered Brier error versus an offset but had higher error than a class-rate constant in 15 SEEG session tests at C = 0.1 under assumed timing
This work supplied the source retrieval and trial alignment lineage for the selected OpenNeuro sessions. We used the inherited source paths, event rules, and fixed signal ranges to check the public bytes again.
usesUnder assumed timing, fixation subtraction showed no clear load probability error reduction with shared features in 15 final SEEG sessions
Its prior absolute power comparison informed the question. Our test changes the target, training sessions, windows, feature selection, and primary score. The cohort is reused.
usesLog band contrasts before averaging canceled separate positive scalar window gains; Brier error did not clearly decrease versus pooling in 15 SEEG tests at C = 0.1 under assumed timing
The public EDF object versions, headers, byte range hashes, and candidate extraction lineage came from this source study through the direct input package. The new source check retrieved all requested ranges again and rebuilt powers for the selected pairs.
usesUnder assumed timing and simulated loss, shaft-loss training at C=1 reduced Brier error in 15 SEEG tests. A rate forecast had less error.
Uses the unchanged archived SEEG feature arrays, waveform audit, source metadata, extraction code, and pinned retrieval receipts. The new analysis tests shaft-loss training and later-session simulated shaft loss.
usesFull correction from eight early trials raised mean later SEEG beta power error in 15 epilepsy patients
Its depth extraction code is in the input lineage. It forms adjacent SEEG contact differences and trial band powers. We preserved the code through the direct parent and rebuilt all used depth features from checked source recordings.
usesSelection of eight SEEG pairs before four-fold validation increased scores for shuffled load labels in 15 patients
We used its verified 30-session power feature files, source metadata, raw interval receipts, and extraction tools for a new contact-selection test. We kept its first and last numbered sessions. We used a different trial filter and three power bands.
usesAdded SEEG power showed no clear letter load gain over scalp EEG and task history across numbered sessions in 15 patients under assumed epoch timing
We reused the selected first and last numbered sessions, 30 SEEG power feature files, source metadata, and EDF extraction code. We retrieved matched scalp EEG files and made a new test of the extra value of SEEG. We also checked one complete SEEG source file and reproduced its feature values.
usesPhase-lag features showed no clear error decrease across sessions for eight-pair SEEG load readers in 15 patients under assumed timing
This study supplies the first and last session cohort, candidate band powers, source metadata, pinned object versions, range hashes, and EDF reader. We retrieve the pinned ranges again to make complex spectra. We also rebuild and check all candidate powers.