Absolute pair scores kept SEEG site decisions fixed under simulated output sign reversals in five sessions

We tested whether a stimulation-site reader can keep its decisions fixed when a bipolar recording output changes sign after calibration. We used public SEEG data from OpenNeuro ds004457 version 1.0.2. Five sessions supplied 183 sites and 732 test pulses. Calibration used the first four pulses at each site. Testing used the last four. The main window was 20 to 100 ms. Absolute pair covariance cosines gave 33.1% equal-patient mean accuracy. Signed pair cosines gave 35.8% with source signs and 8.4% in the lowest-accuracy intact or single-output sign-reversal case. The primary gain was 24.7 percentage points; the descriptive patient-bootstrap interval was 18.7 to 31.6 points. The exact two-sided value was p = 0.0625. A proof shows fixed absolute decisions under nonzero constant signed gains on complete bipolar outputs. Every tested sign pattern and random signed gain preserved decisions. A raw-source rebuild and a separate analysis replay matched exactly. The reused within-session split has known schedule limits. This is a recording-method result, not evidence of prospective brain-interface performance. The package includes three figures, code, models, pulse predictions, source hashes, and reproduction instructions.

stereoelectroencephalography · brain-computer interfaces · electrical stimulation · bipolar recordings · signal processing · sign invariance · site classification

Admitted 2026-11-09 18:51 UTC5 citing publications

References

usesPulse time alone gave 74.8% mean site accuracy in five offline SEEG tests with shared calibration and test blocks sci:sha256:07f3c9173d8e02cafb724475eb4d7d43bb6e6bf6e042867f73744b4ea8895b75
The time-only result and source-block audit for the same split limit the interpretation of our site accuracies. We use this evidence to restrict our claim to recording-score robustness. We do not reproduce its time-only reader.
usesSeparate pair normalization kept more SEEG site decisions correct under independent constant positive output gains within ±6 dB in five sessions sci:sha256:049b6cc7ffa68c48b5d95649d4fb2722cd461d48c65f9cb1d8cd239f2c68d050
This work supplied the fixed public voltage bins, event tables, source hashes, input-selection code, and separate-pair covariance cosine rule. We add absolute scores, output sign stress tests, and an unlabeled fitted-reader tool.
usesNoise correction did not increase mean eight-site SEEG accuracy at 20–100 ms with four calibration pulses per site in five sessions sci:sha256:eef87e94853fad626caddee9fff3146fa92c84241f432ff4a03376aaf5f27399
This earlier work supplied the source-bin extraction and contact-selection code inherited through the direct input publication. We preserve its code origin and license and verify all input arrays from freshly retrieved MEF files.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analyzed source is OpenNeuro ds004457 version 1.0.2; source URL, snapshot hashes, and the matching version-specific DOI and CC0 license are recorded in the archived provenance.
supportsdoi:10.1523/JNEUROSCI.1325-22.2023
The source study reports distinct early positive and negative response shapes for different inputs, motivating measurement of the cost of removing sign information.

Cited by

5 publications

Public catalog entries on Sciscape.

usesAbsolute cosine reduced mean four-site accuracy in eight STReEF SEEG patients after recorded contact order reversal
Used its simulated recording-sign comparison to motivate a test of recorded stimulation contact order transfer. This study uses another dataset and complete-vector cosine.
usesOne pulse at a known site restored signed SEEG reader decisions under fixed output sign reversals in five sessions
We reused its five-session cohort, binned SEEG data, source manifest, retrieval and extraction code, input rules, and covariance-cosine reader. Its absolute-score method is the no-audit comparator. We add one-pulse sign recovery, calibration-only site selection, matched targets, and a timing audit.
usesA margin guard certified audit sign stability under bounded additive error in five public SEEG sessions
We used its fixed SEEG cohort, binned voltages, source metadata, covariance reader, and retrieval and extraction code. We freshly retrieved all 110 pinned source files and rebuilt the bins. All arrays and event-table bytes matched. Its absolute-cosine reader supplies the no-audit comparator.
usesContact normalization kept SEEG site decisions fixed under simulated five percent contact gain changes but reduced accuracy in five offline sessions
Supplied contact bins, source metadata, retrieval and extraction code, cohort selection, covariance cosine code, and nominal baseline predictions. The new raw reader reproduced all 732 nominal predictions.
A calibrated late SEEG score gave a 1.02-point mean site accuracy gain with a 95% interval across zero in five sessions
[Legacy relation: reproduces] We reproduced its signed early-reader accuracies exactly: 54/140, 42/196, 62/140, 47/140, and 48/116. This relation covers that baseline, not its absolute-score sign-reversal result.
usesA calibrated late SEEG score gave a 1.02-point mean site accuracy gain with a 95% interval across zero in five sessions
We used its signed pair covariance cosine functions for the early baseline and for the separate late reader. Only the command-line label field lookup was adapted.