Equal weights for stimulus and choice cells gave 0.511 mean conditional choice AUC in six public SEEG tests at penalty 10
We tested whether equal training weight for four stimulus and choice cells improves choice readout within stimulus direction. We used public processed SEEG high-gamma data from six MetaRDK sessions. Each logistic model used eight channels selected on early trials. At penalty 10, mean conditional AUC on 447 later trials was 0.511 with balanced weights and 0.506 with ordinary weights. The paired gain was 0.0049. A descriptive 95 percent person bootstrap interval was -0.0598 to 0.0681. Three of six people had a gain. These tests did not establish a gain. A stimulus-only control gave 0.741 overall AUC and exactly 0.500 conditional AUC. We preserve a reusable evaluator, all model tests, three figures, source code, compact data, and source retrieval instructions. All six pinned source files were rebuilt and checked. This is an exploratory offline test on a reused cohort. It does not establish online command readout or causal neural control.
brain-computer interfaces · choice decoding · conditional AUC · stereoelectroencephalography · class-balanced training · logistic regression · neural signals
References
4- usesAdding eight SEEG signals to a task model increases offline confidence forecast error on later trials in five of six MetaRDK sessions sci:sha256:f153e0af57732bd1a666747e35778c555a9091f0c980875e17ea5b2460f566ee
We used the source manifest, size-based cohort rule, event-window feature extraction code, and source retrieval code. We downloaded all six pinned source files again and checked the rebuilt features.
- usesEight added SEEG channels increased mean choice log loss on later trials from six MetaRDK sessions at penalty 10 sci:sha256:97e13fb0a793ade1295f735fbb0fc7a8d15bb94c768e56c76e84f4f65ce2055d
We reused the six people, choice label code, fixed early and later trial split, and compact SEEG features. We checked ordinary model probabilities against the preserved parent probabilities. We added weighted training and tests within stimulus direction.
- usesdoi:10.18112/openneuro.ds006253.v1.0.3
The analyzed processed high-gamma files came from MetaRDK, OpenNeuro ds006253 version 1.0.3; the source record declares CC0.
- supportsdoi:10.1038/s41467-025-61744-8
This source paper provides task and signal-processing context for the cited dataset; it does not report the current balancing result.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- Offline eight-channel SEEG models gave 0.484 mean choice AUC within visual directions in 8 new MetaRDK participants
[Legacy relation: reproduces] The six inherited feature hashes matched. All saved ordinary neural-only probabilities were reproduced exactly. This check does not reproduce the earlier cell-weighted model.
- usesOffline eight-channel SEEG models gave 0.484 mean choice AUC within visual directions in 8 new MetaRDK participants
We reused six compact feature files, extraction checks, the neural-only numerical model, and the within-direction AUC method. We added eight full-source extractions and a new eight-person test.
- usesOffline SEEG models increased mean choice AUC by 0.101 after movement within visual directions in six additional MetaRDK participants
We used choice AUC within each visual direction to limit stimulus agreement as an explanation of choice ranking.