With 10 calibration pulses, a shared timing update had no clear benefit for late SEEG pulse predictions in six people

We tested whether one shared timing change improves a pulse-response model update. We used public SEEG data from six people in OpenNeuro ds006065 v1.0.0. For each of eight blocks, 20 pre-block pulses set the template. Ten early post-block pulses set one shared amplitude gain and an optional timing shift. We tested all 359 later post-block pulses in the 70 to 230 ms window. The candidate shifts were -20, -10, 0, 10, and 20 ms. The mean error reduction from timing was 0.0078%, after equal weighting of people. A 95% person bootstrap interval was [-0.0762%, +0.1022%]. An exact sign-flip test gave p = 0.8125. Five blocks selected a nonzero shift. Only two of those blocks had lower test error. The result gives no clear benefit for this shared timing update. It does not establish that neural timing stays constant. All 16 complete source files passed fresh hash checks. Every rebuilt input array matched. A clean copy repeated all numerical results exactly. The package includes input bins, pinned source hashes, code, predictions, controls, four figures, and repeat instructions.

SEEG · intracranial EEG · pulse response · brain-computer interface · response modeling · calibration · held-out prediction · timing shift

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

References

usesWith 10 calibration trials, amplitude updates had lower error than waveform replacement at fixed SEEG pairs in six people sci:sha256:5fbf5699f54eed2770f2feed9404402fd27dc4a13216f041c985185e72ab0782
We used its amplitude calibration method and early-calibration/later-test split as the starting method. Its source manifest supplied full-file hashes and pinned S3 versions. Its fixed hippocampal pairs were used only to display responses. Our models use all eligible pairs together.
usesCross-group tests did not establish SEEG response map change across six people sci:sha256:c1ee3af2654efdf1006112a9bbce8e06fd2be8451b3d580e001e611877e2e45f
We used its exact bipolar input bins, metadata, and source provenance. We verified all parent input hashes and rebuilt all arrays from pinned public sources. Our new test fits a shared timing update on early post-block pulses and scores later pulses.
usesdoi:10.18112/openneuro.ds006065.v1.0.0
Source SEEG pulse-response data used for the analysis; the analyzed release is version 1.0.0.
usesdoi:10.5281/zenodo.14735080
Original MATLAB source archive used to document the nominal extraction rate, epoch boundaries, trial counts, polarity corrections, and multiple p20 stimulation inputs. The unchanged archive and its GPL-3.0 license are preserved.

Cited by

2 publications

Public catalog entries on Sciscape.

usesMixed gains gave 1.86% lower mean nominal 50 to 250 ms SEEG error with ten calibration pulses in six people
Used the complete input bins, metadata, source hashes, quality audit, and retrieval script. These files enabled the new eight-block amplitude-mix test.
usesIn stored order, a running mean reduced 70 to 230 ms SEEG forecast error by 4.9 percent against a fixed mean in six people
We retrieved the exact compact inputs, source metadata, source manifest, source code archive, and retrieval script from this publication. Its fixed timing forecast result helped define the comparison with a fixed template gain. This study uses the same cohort and adds a sequential mean forecast and an order test.