Separate site noise models gave no clear accuracy gain at 20 to 100 ms with four fit pulses in five offline SEEG sessions

We tested whether separate noise models for stimulation sites improve a reader with four fit pulses per site. We used public SEEG from OpenNeuro ds004457 version 1.0.2. The test had 134 sites, 15 recording pairs, and five patient sessions. The primary window was 20 to 100 ms after a pulse. A fixed Gaussian reader mixed 25% site covariance with 75% shared covariance. Mean accuracy was 29.40% with shared covariance and 29.81% with separate site models. The paired change was +0.41 percentage points. The descriptive patient bootstrap interval was -0.68 to +1.50 points. Two patients improved, two worsened, and one had no change. Probability error also gave no clear benefit. These results do not establish equal performance. Calibration and test times overlap across sites, so this is an offline session comparison. We verified 110 raw source hashes. Fresh extraction matched all five archived voltage-bin hashes. A repeat run matched 500 model arrays. The package preserves three figures, voltages, scores, source code, provenance, and repeat steps.

stereoelectroencephalography · SEEG · brain-computer interface · electrical stimulation · site identification · covariance · Gaussian reader · calibration

Admitted 2026-11-09 18:51 UTC2 citing publications

References

usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
This publication supplied the archived 5 ms voltage bins, fixed recording contacts, source metadata, pinned raw hashes, and inherited extraction and read-data code. We rebuilt the inputs from all 110 public source files. The new comparison uses a cosine basis, a restricted 6 mA sample on separate shafts, and separate site covariance with four fit pulses.
usesPulse time alone gave 74.8% mean site accuracy in five offline SEEG tests with shared calibration and test blocks sci:sha256:07f3c9173d8e02cafb724475eb4d7d43bb6e6bf6e042867f73744b4ea8895b75
Its pulse-time control and split audit informed the offline scope. We computed the timing count again in the new restricted sample. Only 32 of 536 test windows followed all fit windows. We did not use its predictions or count these patients as independent evidence.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis used OpenNeuro ds004457 version 1.0.2; the publication records the version and DOI.
usesdoi:10.1523/JNEUROSCI.1325-22.2023
This is the source study for the SEEG data analyzed here.

Cited by

2 publications

Public catalog entries on Sciscape.

usesA limited time trend increased SEEG forecast error by 9.6% against the past mean in five 6 mA sessions at 20 to 100 ms
This package supplied the fixed contact list, voltage bins, event tables, metadata, and raw extraction recipe. We reused the patients and rebuilt the inputs from raw files.
usesLinear maps fitted to pulse responses had more 20 to 100 ms SEEG forecast error than background maps in five selected sessions at ridge strength 0.1
We used its good-contact, non-SOZ, different-shaft and 6 mA site rules, and its nonconstant cosine features for the 20 to 100 ms window. We added a forward-time shared fit and a matched background training task.