Past pulse updates reduced normalized 20–500 ms SEEG forecast error versus fixed five-pulse templates at 86 selected sites in eight epilepsy sessions

We tested whether completed past pulses improve the next SEEG response forecast. The public source was OpenNeuro ds004696 version 1.0.1. We kept 86 readouts chosen with the first five good pulses in eight epilepsy sessions. The main test used all 612 later pulses from 1,042 source good pulses. Each forecast used only past windows completed before the target onset. For the centered 20 to 500 ms window, the scaled all-past mean reduced normalized loss from 0.36882 to 0.35111 versus a fixed five-pulse scaled mean. The mean person percentage reduction was 5.21%, with a descriptive 95% person interval of 3.29% to 7.94%. All eight person differences were positive; exact sign-flip p was 0.0078125. Raw error improved in seven people. The last-five rule had no clear person-level gain. An added exploratory check held 344 final targets fixed. Restoring 182 middle good pulses reduced all-past scaled error by 1.39% on average across people, with interval 0.51% to 2.32%. Earlier studies already tested update rules in this cohort. This is a complete pulse extension and related replication, not a new patient sample. The archive holds exact voltage windows, four figures, all results, a constant-memory forecaster, source hashes, retrieval code, and repeat checks. This offline result does not prove a brain state change or live brain control benefit.

SEEG · electrical stimulation · waveform forecasting · pulse history · epilepsy · offline prediction

Admitted 2026-11-09 18:45 UTC19 citing publications

References

usesFive-pulse noise scaling reduced later single-pulse error for 20–500 ms centered bipolar SEEG in eight patient sessions sci:sha256:fbc06ab469d48eac7a6f54dd97cd24ccd9c7c5ca4b0db4d64340536b3e81428a
Uses the earlier shared-energy scale for mean waveform forecasts. The new code applies that scale to every available past history length.
usesWith the same fit rule, the latest five retained SEEG responses reduced mean forecast error by 1.1 percent against random sets of five at 20 to 500 ms in eight patient sessions sci:sha256:0c97ffc54bebbe1fc1f117cd98063de228096c44c2ccc6a3740218015fc38e51
Uses the earlier restricted pulse archive as the sparse history policy in an added check. Both policies forecast the same 344 later targets. This check restores 182 omitted middle good pulses and does not repeat the random-five comparison.
Recent five-pulse updates with noise scaling reduced mean 20–500 ms centered bipolar SEEG forecast error in eight patient sessions sci:sha256:aab0f09b0cde0fe33fe1115134d13702f7c2609ff81a1be8300ff8d84b1ea5be
[Legacy relation: reproduces] Tests the earlier completed-past update idea in the same cohort. This related replication uses every good pulse at 86 early-selected readouts, rather than first-five and final-five pulses at 886 readouts. It is not an independent patient replication.
usesFive calibration pulses gave mean later signed template AUC of 0.939 compared with 0.859 for one pulse in eight SEEG sessions sci:sha256:1d2aa103275ad7400befec4c899740ede8f2eb59a09897bdc3f4d4e190e54418
Uses its first-five readout choices, candidate pair tables, and exact early and final comparison waveforms. Later forecast scores do not choose the readout.
usesRecorded pulse gaps from 2 to 8 seconds had no clear mean SEEG gain relation in eight sessions with site and order controls sci:sha256:77ca2cd94f79de36a61ae64814637757c4181c890e0e63769f25e428fe358ac5
Uses its 1,042 exact good pulse windows, event joins, source metadata, and retrieval code at 86 fixed readouts. The gap outcome is not an input to the new forecast tests.
usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analyzed processed SEEG derivatives came from OpenNeuro ds004696 version 1.0.1.
supportsdoi:10.1523/JNEUROSCI.2201-22.2023
This is the source experiment paper for the limbic electrical stimulation data used here.
doi:10.1371/journal.pcbi.1011105

Cited by

19 publications

Public catalog entries on Sciscape.

usesStudent t limits scored worse than normal limits for centered SEEG voltage at 86 selected sites after five fit pulses
This work supplied exact all-pulse voltage windows, fixed readout choices, source metadata, and extraction code. We checked 112 copied input files and rebuilt all eight public author derivatives.
usesGeometric medians of five traces increased 20–500 ms SEEG error compared with means in eight sessions without added faults
We used the exact 1042 public SEEG pulse windows, chronological event joins, and fixed readouts. We adapted its source downloader. We reproduced its ordinary and scaled five-pulse mean losses before the new median and fault comparisons.
usesA fixed voltage gate had higher mean SEEG waveform forecast error than matched random selection in eight epilepsy sessions
This archive supplied all 1042 pulse windows, the fixed readout choices, joined events, source hashes, and the public retrieval recipe.
usesSixty block means gave the same SEEG detection AUC as 960 samples on five later pulses at each of 85 selected sites in eight people
The retained range downloader was adapted from this publication. We used it to retrieve all eight complete MATLAB voltage files, check their SHA-256 hashes, and remove each file after exact source comparisons.
usesA fixed adaptive gradient rule showed no clear 20–500 ms SEEG forecast gain in eight sessions
We used its full good-pulse waveforms, fixed early readout choices, source extraction code, and all-history baseline. These inputs made the causal blend test possible.
usesA linear input from voltage size before a pulse, fit in other patients, increased SEEG forecast error at 86 readouts in eight epilepsy sessions
Uses its exact 1,042 pulse windows, event joins, source metadata, and retrieval code at 86 fixed readouts. The new test forecasts log voltage size with an input from before the current pulse. It does not compare full waveform losses with size losses.
usesA linear pulse-gap rule did not reduce mean SEEG size forecast error at 85 sites in eight epilepsy sessions
This work supplied the exact 1042 selected good-pulse windows, pulse history, and GPL source downloader. We used them for the early-only forecast split and fresh source comparison.
usesA test of 610 later pulses at 86 sites did not establish smaller adjacent 20 to 500 ms SEEG differences in eight patients
Uses all 1,042 exact good-pulse waveform windows, frozen first-five readout choices, source event mappings, and retrieval provenance. The new order audit tests 610 guarded later pulses at the same 86 sites. It is not a new patient cohort.
usesAn updated local RMS model did not establish lower SEEG forecast error at 86 selected readouts in eight sessions
We used its exact pulse mapping, fixed readout tables, complete-source downloader, and source hashes. We checked every inherited before-pulse and after-pulse waveform against its archive. The new study measures different 400 ms windows and updates a separate local slope.
usesFive later pulses had 6.5 percent lower SEEG sample variation rank after stimulation than before it at 85 selected sites in eight people
Its inherited range downloader checks full source sizes and hashes. We use this code for fresh extraction checks of all eight complete public MATLAB source files.
usesSEEG bands with fitted shapes had 6.8% lower normalized half-width than constant bands on average for 20–500 ms responses at 86 sites in eight patients
This publication supplied the exact 1,042 retained voltage windows and their pulse tables. Its retrieval code rebuilt all eight processed public source files.
usesSeven calibration pulses gave a 3.33% mean person error reduction over five for centered 20 to 500 ms SEEG at 86 selected readouts in eight sessions
We use its complete chronological pulse windows, fixed readouts, public source recipe, and source records. We add fixed counts of five and seven with the same final three targets and exact target accounting.
usesF limits covered 89.8 percent of later three-coefficient SEEG responses after five fit pulses at 86 selected sites in eight patients
Its inherited range downloader checks pinned source sizes and hashes. The recipe holds one derivative file at a time.
usesA 41-sample variance mean reduced later normal SEEG voltage log loss with five fit pulses at 86 selected sites in eight patients
Used its range download and full source hash code through the direct input publication. That code enabled fresh checks for all eight author derivative files.
usesCalibration across seven checks reduced before-pulse SEEG calls at 85 selected sites in eight people
We used the adapted range download code to retrieve complete pinned MATLAB files one at a time and check every saved waveform.
usesA variance rule showed no clear 20–500 ms SEEG forecast gain at an equal pulse budget in eight offline sessions
Its earlier complete pulse extraction and GPL downloader are code and data ancestors of the immediate input. We used that downloader through the input archive and preserved its source and license.
usesAn AR(16) model with a 1 percent floor gave no clear 20–500 ms SEEG detection gain after five calibration pulses in eight epilepsy sessions
This publication supplied the exact all-pulse voltage windows, event joins, fixed readout choices, source metadata, retrieval code, and inherited source audit. We kept its early-selected readouts and tested a new background detector on every later good pulse.
usesA fixed one-pair model raised mean relative error by 12.4% for later five-pulse SEEG forecasts in seven people
This study uses the pinned OpenNeuro source records, clinical metadata, upstream preprocessing records, and GPL download code. Fresh source checks rebuild 190 mean traces for 95 site-output pairs.
usesAn AR(16) model with a 1 percent floor gave no clear 20–500 ms SEEG detection gain after five calibration pulses in eight epilepsy sessions
This publication supplied the exact all-pulse voltage windows, event joins, fixed readout choices, source metadata, retrieval code, and inherited source audit. We kept its early-selected readouts and tested a new background detector on every later good pulse.