Offline SEEG power before a response separated task and rest across cue durations in six patients

We tested whether a change in visual cue duration reduced task-versus-rest decoding in public SEEG. The main test used 754 selected trials from six patients with epilepsy in OpenNeuro ds004473 version 1.0.1. A fixed model used power from the 500 ms before a response. Each model trained on one recording block and tested on other blocks. Mean ROC AUC was 0.725 in the other cue condition and 0.741 in the same condition. The mean change was -0.016, with a descriptive 95% patient bootstrap interval from -0.034 to 0.001. The paired sign-flip p value was 0.1875. The rest control reached AUC 0.510. Training on the first block reached other-condition AUC 0.727. These results support offline task-versus-rest decoding across the tested cue conditions. They do not establish equal performance or motor-intent specificity. The package preserves features for 849 trials from seven patients, six figures, all test scores, source code, metadata, exact source object versions, byte-range hashes, and reproduction instructions. Complete raw-file hashes were recorded but not verified.

stereo-EEG · task-versus-rest decoding · cue duration · brain-computer interface · recording-block transfer · ridge regression

Admitted 2026-11-09 12:26 UTC3 citing publications

References

supportsA fixed SEEG decoder separated windows before keypress from rest on later trials in eight patients sci:sha256:bd6c59023a95ca7dc23d6549123fb8e0b3724d75412db49c1e004ce64fc1dd57
A separate analysis of the same source recordings also separated windows before a response from rest. Our six-patient fixed model supports that narrow offline result and adds a same-condition versus other-condition block comparison. The patient cohorts overlap. The models, source version, trial subsets, and metrics differ, so this is not an independent-cohort replication or a direct score comparison.
usesSEEG power before a response gives 57.7% mean balanced accuracy for two cued keys on later trials in eight recordings sci:sha256:9fcbc6aad0f73578b656a03c940b9381ab5791e8f9875ee291d7fc423328ee2d
The source correctness audit identified a key and correctness-flag mismatch in patient 1. We used that finding to add a metadata audit after the main analysis. We confirmed 102 synchronized, source-kept, matching-key left trials within the deadline had a zero correctness flag. The six primary patients had no such mismatch. The planned trial filter was retained.
usesdoi:10.18112/openneuro.ds004473.v1.0.1
The analysis used public SEEG signal ranges and source task events, behavior records, channel flags, and trial keep masks from OpenNeuro ds004473 version 1.0.1.
doi:10.1088/1741-2552%2Facae0a

Cited by

3 publications

Public catalog entries on Sciscape.

usesSame-trial power subtraction did not show a clear Brier error gain for eight SEEG pairs on later blocks in six recordings
Its task-versus-rest transfer test across two cue conditions informed the use of separate readers for later blocks in the same condition. Our target is the recorded response key. We did not copy its feature subset or fitted readers.
usesSEEG key decoding from the first fast block had a mean AUC of 0.54 in later slow blocks in six patients
Its task-state transfer study supplied the cue-duration interpretation and complete-block comparison used to choose the present design. We tested recorded-key identity on a different version and feature pipeline. The results concern reused recordings and are not an independent replication.
usesA 500-sample rest reference gave lower mean later-block SEEG balanced accuracy with 32 calibration trials in six patients
This publication supplied the exact SEEG power arrays, trial selection records, fixed source metadata, EDF headers, pinned S3 range hashes, and extraction code used to test a separate rest power reference on strictly later blocks. All 849 feature trials were rebuilt and matched exactly. The recordings are the same cohort.