Eight selected sites raised accuracy in all five SEEG sessions but raised log score in four

We tested a calibrated template reader in five public SEEG sessions from OpenNeuro ds004457 version 1.0.2. Three early pulses per site fitted the templates and selected eight sites. A fourth pulse fitted one probability scale for all site-set sizes. Four later pulses per site supplied the tests. The full sets had 29 to 49 sites. Mean accuracy was 97.5% for eight sites and 38.2% for all sites. The probability log score was 2.908 and 2.171 bits per pulse, respectively. The paired mean score change was +0.737 bits per pulse. The score rose in four patients and fell in one. The descriptive patient bootstrap interval was [0.058, 1.273]. This score includes the loss of label entropy when sites are removed. Its population expectation is an information lower bound under a uniform site prior and a fixed reader. The finite test score is not a certified information bound or brain-writing capacity. The sessions and test pulses were reused from prior work. The split was later within each site, but was not globally forward in time. The package preserves code, compact voltage data, source hashes, probability tables, controls, and figures.

stereoelectroencephalography · electrical stimulation · probability log score · site selection · neural response decoding

Admitted 2026-11-09 17:56 UTC10 citing publications

References

usesA joint SEEG shape decoder did not show a clear mean gain over one selected pair in four patient sessions sci:sha256:81f77d4f52365d5b14fed7d3d6db7b430a2cf478b9a6a6062540c4e90bbd8e7c
Uses extraction code and read-data logic inherited through the cited covariance study. The package preserves the source license and code ancestry.
usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
Uses the exact voltage bins, source metadata, contact eligibility rules, and fixed covariance rule. A fresh raw extraction gave identical input arrays. This study adds probability calibration and site-count log scores.
usesFour calibration pulses per site selected eight sites with higher mean later-pulse identification accuracy than random sets in five SEEG sessions sci:sha256:734ed418ba7748370955abc8656549ca19d02417622be27efe64bff0eec4d5eb
Uses the prior eight-site accuracy result to choose the question and the primary site count. The new score includes label entropy and probability log loss. The patients and later test pulses were reused.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The study analyzes public SEEG recordings from this exact OpenNeuro dataset version.
usesdoi:10.1523/JNEUROSCI.1325-22.2023
This paper describes the source experiment for the analyzed stimulation recordings.

Cited by

10 publications

Public catalog entries on Sciscape.

usesA scale with template error reduced log loss against T/4 for four averaged SEEG responses in five patients
This study reused the voltage means, source metadata, pulse split, site selection, and single-response reader. It added fixed scale comparisons for repeated responses and reproduced the original single-response scores.
usesDivision of evidence by four reduced mean log loss with a fixed Gaussian reader in five offline SEEG sessions
Uses exact voltage inputs, pulse roles, eligibility rules, and the single-pulse probability reader.
usesMeans of two responses lowered log score per pulse for calibrated site readers in five SEEG sessions at 20 to 100 ms
We used its compact voltage bins, source metadata, and fitted-reader definitions. We used a new nine-pulse rule, two calibration pulses, and a one-response versus two-response pulse-cost comparison.
usesSite templates increased 20–100 ms shaft log scores over shaft means in five SEEG sessions
This work supplied the exact 5 ms voltage inputs, event tables, source metadata, raw file manifest, and extraction code. Their hashes were checked before the new shaft analysis.
usesSite variation scales give mixed probability gains in five SEEG sessions with three template pulses and separate calibration
We used its three-template and fourth-pulse probability-fit split, fixed one-percent uniform probability mixture, and held-out log-score procedure. We kept all eligible sites and tested variation scales. We did not use its selected site sets.
usesCoordinate spacing gave lower mean eight-site accuracy than random sets in five offline SEEG tests at equal calibration pulse counts
Uses its response-based eight-site selection result to define the new calibration-cost question. This study selects sites from coordinates and refits each reader with only the selected calibration pulses. It does not claim a matched comparison with the prior probability reader.
usesOne log RMS value per pair reduced site accuracy in five SEEG sessions with a fixed reader
This package supplied the source files and provenance records. We adapted its retrieval, extraction, and input-reading code. Those tools enabled the raw rebuild and the matched feature comparison.
usesSelection from three windows and one early window gave the same mean lowest site accuracy in five offline SEEG sessions
We extended its greedy site selection rule. The new rule uses the minimum fitted pair distance across three response windows to select eight sites. The earlier report uses the same cohort. This work is an exploratory extension, not independent confirmation.
usesFinite template scaling increased pulse use without reducing site error at a 90% threshold in five offline SEEG tests
The site-bank selection and single-response probability methods are inherited through the input package and reader code.
usesTwo calibration pulses gave a mean score gain of 0.062 bits over eight sites for 20 to 100 ms SEEG responses in five sessions
This work supplied the probability-score framework, one-percent uniform probability mixture, and fixed eight-site comparison. We extended the method with two separate size-choice pulses. The source cohort and final test pulses were reused.