Subtracting the median voltage of five other contacts reduced source response label AUC estimates at 10–90 ms in two SEEG recordings

We tested whether subtracting the median voltage of five other contacts improves response label ranking. We used public SEEG data from OpenNeuro ds004774, release 1.0.0. Two recordings supplied 630 pulses, 378 site/contact cells, and 566 expert judgments. The score used the largest absolute ten-pulse mean voltage from 10 to 90 ms, divided by baseline standard deviation. Source AUC estimates were 0.905, 0.767, and 0.892. Median subtraction gave 0.833, 0.744, and 0.837. Paired AUC differences were -0.072, -0.023, and -0.055. Only the first primary 95% site-resampling interval excluded zero. Mean subtraction also reduced all three estimates. Response peaks fell, while baseline noise changes differed between recordings. A separate 20-90 ms check was added after the primary results and gave the same direction. The experts judged source-reference waveforms. These results measure transfer to those labels and do not identify neural ground truth. The package preserves data, source versions and hashes, code, tables, four figures, and reproduction instructions.

stereoelectroencephalography · electrical stimulation · evoked responses · signal referencing · area under the ROC curve · expert labels

Admitted 2026-11-09 18:31 UTC2 citing publications

References

usesPeak confirmation had lower response label AUC point estimates than a peak from ten pulses at 10–90 ms in two SEEG recordings sci:sha256:99e496cb85e0cce73dd6f273e89a55d7e07b99da7a9cc458130a30766d18b9ce
We used its unchanged native SEEG epochs, expert depth labels, source metadata, event selection, byte-range provenance, and raw retrieval code. We also used its source-reference scores and AUC estimates to check the new reference comparison.
usesdoi:10.18112/openneuro.ds004774.v1.0.0
Source epochs, labels, metadata, and event records are from OpenNeuro ds004774 release 1.0.0.
supportsdoi:10.1016/j.jneumeth.2025.110389
Describes the ER-detect dataset and response detection methods used as context; this study is not a full pipeline reproduction.
supportsdoi:10.1016/j.jneumeth.2024.110153
Prior work cited for response bias in common average references and the CARLA channel-selection approach.
supportsdoi:10.1101/2024.01.09.574915
Source for the ER-detect baseline deviation threshold context and visual annotation method; this study uses a simplified score.

Cited by

2 publications

Public catalog entries on Sciscape.

Using N1 labels reversed the mean reference AUC point difference for expert SB at 10 to 90 ms in UMCU21 SEEG
[Legacy relation: reproduces] All six source and equal-weight mean AUC point values with combined labels match exactly. Both studies use the same recordings. This check adds no independent patients. It does not reproduce the other-five-contact median test. We add a later check with the source N1 label definition.
usesRemoval of shared voltage reduced site identification accuracy for a fixed method in two SEEG recordings with six contacts
The earlier reference analysis on these recordings supplied the comparison with response labels and shared energy in ten-pulse means. We used that scope to define a separate test of repeated site differences and later-pulse site identification.