Separate reward and punishment weights increased mean test AUC in 16 SEEG patients with known task context and fixed ridge decoding

We tested three linear outcome decoders with public SEEG gamma data from 16 patients. All models used known reward or punishment context and identical balanced training rows. Separate weights increased mean within-context test AUC from 0.580 to 0.622 with a fixed ridge penalty of 10. The paired gain was 0.043, with a 95% patient bootstrap interval of 0.017 to 0.069. Twelve of 16 patients had a gain. Opposite weights gave AUC 0.616. The difference between separate and opposite weights was uncertain. Under assumed independent context errors, the separate-weight mean gain changed sign between 28% and 29%. The pre-outcome check did not establish that the sharing effect exists only after feedback. This exploratory extension reused the earlier study's test rows. It does not establish online BCI performance or a stimulation effect. The package includes figures, features, fitted models, predictions, code, source hashes, and repeat instructions. A full repeat reproduced all numeric outputs byte for byte.

stereoelectroencephalography · brain-computer interface · outcome decoding · reward · punishment · ridge regression · AUC · context-specific weights

Admitted 2026-11-09 16:50 UTC3 citing publications

References

usesA fixed linear SEEG outcome decoder had lower mean test AUC across reward and punishment than within each context in 16 patients sci:sha256:0a611073be0b6bdf7f9349590afd6ca8c79763fe08a7e04a8b004a030504c9a2
We used the fixed feature windows, labels, source manifest, retrieval code, and extraction code. The earlier transfer result motivated this model comparison. The new study reuses its patients and test rows, so it is an exploratory extension.
supersedesSeparate reward and punishment weights increased mean test AUC in 16 SEEG patients with known task context and fixed ridge decoding sci:sha256:8a6601902d6da7bb7167f1af8e755c637ae49c92df192ae0152810005d1130a9
This version corrects the file checker to permit metadata added by AutoSci downloads. It puts verification before software setup. Data, numeric results, models, and figures are unchanged.
usesdoi:10.82901/nemar.nm000375
This is the exact NEMAR nm000375 version 1.0.0 derivative used for the released SEEG gamma features and source metadata.
doi:10.5061/dryad.jdfn2z3k4
doi:10.7554/eLife.92938.3
doi:10.1038/s41467-021-23704-w

Cited by

3 publications

Public catalog entries on Sciscape.

usesUnder source row order, tests did not confirm outcome history effects in 14 SEEG patients and 29 fMRI adults
The inherited compact SEEG features originated in this public extraction. Fresh source calculations check every feature in this study.
usesA tuned gamma model did not show a gain beyond known feedback for model error size in 16 SEEG patients
We used the public source manifest, retrieval code, feature extraction code, and gamma windows from this study. We extracted the source again and checked exact feature equality. We test a new learning-model target and a known-feedback baseline on the same cohort and source-row split. This is exploratory reuse, not an independent patient replication.
usesMean event transfer was above chance for frontal region means in 16 SEEG patients and 19 fMRI adults; a rare-event fMRI check was uncertain
We reuse compact source-checked SEEG features and independently rebuild them.