Graph fits from past pulses reduced mean squared log error by 30 percent for reverse SEEG forecasts at 20 to 500 ms in six reused sessions

We tested a graph correction for a missing SEEG response-map entry. All fitted pulse windows ended before the first test window started. The test used 167 reciprocal pairs in six reused epilepsy sessions from OpenNeuro ds004696, version 1.0.1. The size was the RMS of a five-pulse mean at 20 to 500 ms. Each patient had equal weight. Mean squared log error fell from 1.943 for the reverse forecast to 1.353 for the past graph with a reverse fallback. This was a 30.4 percent reduction. Four of six patients improved. The mean gain was 0.590, with a descriptive 95 percent patient bootstrap interval of 0.099 to 1.053. The exact patient sign-flip reference was p = 0.1875. Past paths existed for 154 of 167 pairs. The past rule kept 49.1 percent of the known full-session gain. Direct calibration had lower error in each patient. A similar gain occurred before pulses. The result therefore supports a bounded voltage-map benchmark. It does not establish a stimulation-specific neural effect or a population benefit. Every connected fit used other readouts from the target site. Source quality labels and the cohort remain retrospective. The package includes exact pulse products, all forecasts, timing and graph audits, a connection proof, code, three figures, and source recipes. A fresh pinned public source for patient 06 passed all compact-product checks. A second analysis run produced eight byte-identical result files.

SEEG · electrical stimulation · response maps · graph least squares · voltage forecasting · time-restricted calibration

Admitted 2026-11-09 19:50 UTC2 citing publications

References

usesFive pulses per site gave higher later 20–500 ms voltage scores than the nearest midpoint rule in seven offline SEEG sessions sci:sha256:5a3b0ef7c60d9ccb6ab72780104c08522d5c00ef2ac425835fefc99751512412
This publication is the immediate data parent of the unchanged all-pulse Gram products. These products supply exact five-pulse mean sizes and pulse row maps. We keep their file hashes and source receipts.
usesCycle components of SEEG pulse size ratios were invariant to fixed scalar site gains in six sessions sci:sha256:730f1c79a2cacd4597fa7f3961c4db1c917d7c75ca8bf8bd07342218f7993c63
This publication supplies the exact-contact graph definition and the equal-weight node-ratio model. We use those definitions with a time cutoff and an explicit rule for disconnected endpoints.
usesA graph correction reduced reverse voltage size error in six offline SEEG maps at 20 to 500 ms sci:sha256:96389bb770bbad66c77753b1c60eff6f8a53dd005959a2d778bb7e5d0b5a68f6
This package supplies the known full-session graph benchmark, the exact compact inputs, and the fixed 167-pair cohort. We reproduce all 668 window forecasts as a code check. We add a strict past-window rule, a connection check, and a reverse fallback. The patient sample is reused.
usesA geometric median reduced 20–500 ms SEEG error with one simulated gain error among five calibration pulses in eight patients sci:sha256:c37805e37f6922e0de8e20f0231c72f98687698645324f2c0c6261e9d3ce9010
This publication is the origin of the inherited all-pulse products and source retrieval recipe. We use its fixed source selection and Gram definitions. We verify patient 06 from the pinned public derivative.
usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analysis uses compact pulse products derived from OpenNeuro ds004696, version 1.0.1.
doi:10.1523/JNEUROSCI.2201-22.2023
Describes the acquisition and limbic response study for the source SEEG recordings.

Cited by

2 publications

Public catalog entries on Sciscape.

usesUniform graph fits reached the smallest bound for one ratio error among unit flows in 105 of 154 connected SEEG pairs across six sessions
This work supplied the public pulse products, 167-pair cohort, past-window cutoff, size loader, and uniform graph forecast code. We retained those choices and reproduced all 668 uniform pair-window forecasts before testing the graph cut rule.
usesA fixed covariance fit had 6.6 percent more mean squared log error than uniform fits at 20 to 500 ms in six reused SEEG sessions
Used the exact all-pulse products, past donor graph rules, and uniform forecast code. The uniform forecasts were reproduced before the new covariance rules were compared.