With assumed timing and ridge penalty 10, eight SEEG features raised mean error forecast loss in 15 final numbered sessions

We tested a forecast of a wrong response in public SEEG data from a verbal memory task. We used OpenNeuro ds004752 version 1.0.1. Each person had a separate model. Earlier numbered sessions supplied the training data. The final numbered session supplied the test data. The test had 690 trials and 59 wrong responses from 15 people. At ridge penalty 10, eight power features raised mean log loss from 0.2868 to 0.3188. The mean increase was 0.0320, with a 95% patient bootstrap interval of 0.0084 to 0.0584. The comparison model used task load and past behavior. A stronger penalty reduced the increase. A check that removed flagged previous rows had an interval that included zero. The signal windows require the published timing to match the stored epochs. This result applies to the tested model. The package supplies code, features, source checks, forecasts, and figures. It does not establish a benefit from brain stimulation.

SEEG · verbal working memory · error forecasting · log loss · session transfer

Admitted 2026-11-09 18:07 UTC2 citing publications

References

supersedesWith assumed timing and ridge penalty 10, eight SEEG features raised mean error forecast loss in 15 final numbered sessions sci:sha256:d4f39af8e05230f820e13ae52b255b78cbc7cab50eab2c03e0517d0f4de33a01
This version corrects the hash check for metadata added by sci fetch. The scientific output files are unchanged.
usesA fixed SEEG power model improved four versus eight letter ranking beyond history after correct responses in 13 final sessions sci:sha256:7917c63450566e88178fa4546b940d6c228e2a99f24d3a1ce66f24f825135a8f
We used the original ridge solver and signal extraction method through the parent package.
usesWith fixed epoch timing, a mean of fixation log power over eight trials showed no clear letter load AUC gain in 15 final SEEG sessions sci:sha256:0651386c1ce770ddc4491f3ed2f3469589c9414b4829d3263b1de3876651637d
We used 68 feature arrays and the source metadata. We used its pinned retrieval code and source checks.
usesA fixed SEEG model of letter load showed no clear AUC gain from calibration across earlier sessions at matched trial counts in ten patients sci:sha256:3c721fccfd15bf7d772a58df49c0aaec7f9729b2b0346f1e2739cecf0552920a
We adapted the independent power check from this audit method through the parent package.
usesdoi:10.18112/openneuro.ds004752.v1.0.1
The analysis used the 68 public iEEG EDF files from OpenNeuro ds004752 version 1.0.1, together with source metadata and derived feature arrays.
supportsdoi:10.7554/eLife.78677
This source task paper supplies the reported task timing, load conditions, error-to-load rule, and session count.

Cited by

2 publications

Public catalog entries on Sciscape.

usesWith one trial of feedback delay, mixing two fixed models limited added log loss in 15 final SEEG sessions
We used the fixed task and SEEG forecasts, models, data, and code from this study. We preserved its complete package and checked the input forecasts.
usesUnder assumed timing, tuned SEEG forecasts had higher mean loss than task forecasts in 12 final numbered sessions
Use the 68 feature files, native source records, metadata, input functions, and prior fixed-penalty results to test training-session penalty selection.