At rate 1, learned weights reduced 20 to 500 ms SEEG forecast error against equal weights but raised error against scaled means in eight recordings

We tested weights learned from earlier SEEG responses in eight people with epilepsy. The public source was OpenNeuro ds004696 version 1.0.1. The primary target was the next included good pulse at the same site. Its voltage window was 20 to 500 ms. At rate 1, learned weights reduced error by 1.50% against equal weights on the same four forecasts. The 95% person bootstrap interval was 0.63% to 2.43%. Error was lower in all eight people. However, the mixture increased error by 0.40% against the single scaled past mean. This increase occurred in all eight people. The study included 85 sites, 879 readout streams, and 562 distinct target pulses. We provide exact Gram inputs, four figures, an error bound proof, and source retrieval records. Independent waveform calculations checked all scores. A fresh full source file for sub-05 reproduced 68 Gram matrices exactly. All 44 metadata files passed fresh hash checks. The other seven full source extractions were inherited. This was a retrospective study of good pulses in a reused cohort. The result does not establish behavior decoding or stimulation control performance.

SEEG · sequential forecasting · exponential weighting · electrical stimulation · neural signals · waveform prediction · forecast error · epilepsy · OpenNeuro

Admitted 2026-11-09 20:03 UTC10 citing publications

References

usesA noise-scaled linear time trend increased SEEG forecast error for the next pulse at 20 to 500 ms in eight patients with epilepsy sci:sha256:9adb70e1c0e233690091510e00f8b9429a610943970f5417cd5d58644c8139e2
We used the exact all-pulse Gram inputs, native voltage example, extraction helpers, and source records. We calculated new sequential mixture forecasts and errors. We rebuilt sub-05 from its complete public source and checked all metadata files again.
usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analysis used the exact all-pulse Gram inputs derived from OpenNeuro ds004696 version 1.0.1.
usesdoi:10.1523/JNEUROSCI.2201-22.2023
This is the source study credited by the dataset for the human limbic stimulation recordings used here.

Cited by

10 publications

Public catalog entries on Sciscape.

usesRemoving one of five pulses before noise scaling raised 20 to 500 ms SEEG forecast error on recorded data but cut error with one simulated gain 4 fault in eight people
We used its all-pulse Gram matrices, native waveform examples, source metadata, and retrieval code to fit and score fixed forecasts.
usesNested pair selection did not show lower absolute score error at 83 SEEG sites in eight recordings
We copied its source retrieval and extraction helpers from the descendant package. The helpers define contact pairs, regions, source-good trials, and waveform windows.
usesA fixed adaptive gradient rule showed no clear 20–500 ms SEEG forecast gain in eight sessions
Its four-expert weight result motivated the strong scaled all-history comparator. We tested a continuous two-forecast blend with a different update rule and readout set.
usesRemoving one of five pulses before noise scaling raised 20 to 500 ms SEEG forecast error on recorded data but cut error with one simulated gain 4 fault in eight people
We used its all-pulse Gram matrices, native waveform examples, source metadata, and retrieval code to fit and score fixed forecasts.
usesThe final five pulses showed higher mean SEEG alignment after stimulation across eight patients
We used its all-pulse Gram inputs, native voltage example, source records, and extraction helpers. We calculated new pulse-pattern scores across different recording shafts. A fresh complete source file for sub-05 checked all 68 window matrices.
usesSelection on two shafts increased SEEG shape retention after simulated worst shaft loss at 82 sites in eight recordings
This publication supplied the exact all-pulse Gram archives, native voltage example, source metadata, retrieval helpers, source hashes, and inherited audit records. This study tests a new four-contact shaft-loss shape endpoint on the same cohort.
usesAt selected pairs in eight SEEG recordings, scoring all later pulses changed normalized 20 to 500 ms error of a fixed five-pulse forecast by less than 0.7% per person
We used copies of the source metadata and their dataset release record to define the cohort. We checked every copy against the pinned source tree.
usesA flexible calibration rule reduced 20 to 500 ms SEEG forecast error against concentrated calibration at the same budget in eight recordings
We used its all-pulse Gram data, native voltage example, source metadata, pinned source manifest, and retrieval helpers. These inputs enabled exact loss curves for two to five calibration responses. We freshly rebuilt sub-01 and checked the copied input hashes against its archive.
usesOffline response noise allocation had an uncertain 20 to 500 ms centered bipolar SEEG forecast gain at three pulses per site in eight recordings
We used the GPL retrieval helpers and source-audit structure that the direct data parent copied from this work. These files support the pinned public-source rebuild and the event-to-trial mapping.
usesA fixed pulse limit increased detection of imposed SEEG gain trends in eight recordings under an exchangeable pulse model
We used its GPL bipolar extraction helper through the direct source publication. It supplies the window definitions, contact mapping, and source event selection.