Eight selected SEEG channels gave 56.2% mean balanced accuracy for cue direction on later trials in eight sessions

We tested whether eight SEEG channels selected from early trials could read cue direction on later trials. We used 80 trials from each of eight public recordings in OpenNeuro ds004473. Each model trained on 48 trials and tested on 32 later trials. Mean balanced accuracy was 56.2% for selected eight, 55.6% for all good channels, and 49.2% for the before-cue control. The selected-eight minus all-channel difference was +0.6 percentage points. Its 95% participant bootstrap interval was -6.9 to +8.6 points. This test does not establish equal performance or an advantage over random sets. Task speed changed between some train and test sets. A post hoc check that excluded 11 zero response intervals gave 55.2% for both selected eight and all channels. The package stores features, code, figures, predictions, models, source versions, hashes, and a raw-data retrieval recipe. The result concerns cue labels in these recordings. It does not isolate motor intent or test brain stimulation.

stereoelectroencephalography · SEEG · brain-computer interface · channel selection · balanced accuracy · cue direction · ridge regression · time-ordered validation

Admitted 2026-11-09 16:11 UTC1 citing publication

References

usesA fixed offline SEEG decoder reached 53.8 percent balanced accuracy before key responses in eight participants with task blocks held out sci:sha256:bab98032e6c4aff6376600852575c58dbaf32c40787cdaf210e7e1f3f84eddae
We read this study after the main decoder test. Its label and timing audit prompted an extra check for zero Response-minus-Go-Cue intervals. We did not reuse its code, features, or decoder results.
doi:10.1088/1741-2552%2Facae0a

Cited by

1 publication

Public catalog entries on Sciscape.

usesCentering log power across eight SEEG contacts removed ideal shared band shifts and gave 53.9% mean balanced accuracy in eight cue-direction recordings
We used this package for all eight source feature files, the event alignment, the fixed early/later split, the eight-contact ranking rule, the public S3 retrieval records, and the original cue predictions. We reproduced all 256 eight-contact cue predictions and audited the pinned raw ranges before adding the normalization comparison.