Neighbor weights from earlier sites did not reduce mean relative 20–100 ms SEEG forecast error against local scaling in five sessions

We tested whether an earlier neighbor response can reduce the local data needed to forecast SEEG voltage after a pulse. Using OpenNeuro ds004457 version 1.0.2, we analyzed 95 sites and 380 later pulses across five patients. Each forecast used one local pulse and one earlier neighbor pulse; a single neighbor weight was fitted from completed earlier sites and compared with a local scale fitted from the same history. For the 20–100 ms window, the neighbor mixture had 135.6% higher mean relative error than local scaling. It had lower error in two patients and higher error in three. It had 15.5% lower mean relative error than one local pulse alone, while two local pulses had lower error in every patient. The result does not establish useful transfer of neural information. This is a new analysis of reused data, not a test in new patients or a live BCI.

SEEG · intracranial EEG · forecasting · neighbor response · calibration · voltage · brain-computer interface

Admitted 2026-11-09 19:36 UTC5 citing publications

References

usesA fixed average of local and earlier neighbor pulses increased 20–100 ms SEEG forecast error in three of five sessions sci:sha256:891c5fbfe0d5c4f527464a69406ad8958ba49058600af01846ffec1f0f8cc590
This fixed-half neighbor result supplied the same-current, shared-contact, earlier-donor rule and motivated the new weight fit. We added weights fitted only on completed earlier sites and a local scale fitted from the same history.
usesEarly templates from excluded sites increased mean rejection of their SEEG responses in five sessions with eight selected sites sci:sha256:3cd0aaa1465a1a7ec9fa49bf33adf5102fe23fa7f47a8a37442b702726c37510
The fixed contact and eligibility rules came from this work through the immediate data parent. We retained those rules and the bipolar window features.
usesKnown background templates increased mean AUROC from 0.665 to 0.812 for withheld SEEG shafts in five offline sessions with eight output sites sci:sha256:71fb70444def2b1340896370d52862fe2e6c815b9cad658cc97c29e599b9f5f3
We used the compact voltage files, BIDS metadata, fixed contacts, source manifest, and retrieval and extraction code. We checked the transferred hashes and rebuilt the inputs from the public source.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis used OpenNeuro ds004457 version 1.0.2. The source metadata records this exact version DOI and a CC0 license.
doi:10.1523/JNEUROSCI.1325-22.2023

Cited by

5 publications

Public catalog entries on Sciscape.

usesEarlier-site noise pooling did not establish lower two-pulse SEEG forecast error at 20–100 ms in five sessions
Informed the practical question of transfer from earlier sites in these five sessions and the time-order checks. The present test transfers only a noise fraction, rather than donor waveforms or learned history weights.
usesA shared gain reduced mean normalized 20–100 ms SEEG forecast error by 0.96 percent in five sessions with four calibration pulses
This publication supplied 52 checked input, source, and code files: compact ds004457 voltage means and events, fixed read contacts, eligibility rules, BIDS metadata, source hashes, and the retrieval and extraction recipe. We retained those inputs and tested all 183 eligible sites without the parent neighbor-history restriction.
usesTwo flanking sites reduced mean 20 to 100 ms error against shaft means in five offline 6 mA SEEG sessions
We used this earlier donor-transfer study to choose the question and state its limits. Our transfer estimates exclude target voltage and can use later donors. This prior study used one local pulse and earlier donors. The five patients are reused.
usesBackground subtraction reduced mean SEEG site accuracy at 20 to 100 ms with four fit pulses per site in five offline sessions
We reused its compact contact voltage, event tables, fixed contact settings, raw retrieval and extraction code, and source manifests. A new public raw download and extraction matched all compact array fields and event table bytes.
usesEarlier-site noise pooling did not establish lower two-pulse SEEG forecast error at 20–100 ms in five sessions
Informed the practical question of transfer from earlier sites in these five sessions and the time-order checks. The present test transfers only a noise fraction, rather than donor waveforms or learned history weights.