Median block powers gave higher task-state accuracy than mean powers for simulated 1 mV, 10-sample SEEG faults on one contact in eight recordings

This study used fixed public SEEG inputs from OpenNeuro ds004473 version 1.0.2. It compared the mean and median of five separate 100-sample band powers. Each decoder used the same eight bipolar pairs and 500-sample task window. Early trials supplied the fits. The later test had 674 trials from eight epilepsy recordings. The primary digital fault changed one source contact for 10 samples, with opposite 1 mV values in its two halves. Every selected contact, all 491 starts, and both signs were tested. Mean balanced accuracy was 56.85%; median accuracy was 67.79%. The paired gain was 10.94 percentage points, with a 95% person bootstrap interval of 5.97 to 16.44. All eight recordings showed a gain. On unmodified inputs, the median lost 4.10 points relative to the mean. An amplitude-independent outer score bound fixed the decision on 5.82% of later windows for arbitrary finite changes in any 10 consecutive samples on one contact. The bound assumes separate local processing and an available recorded reference. Exact inputs, source hashes, code, fitted models, results, a proof, and four figures are preserved. Direct voltage checks, independent library checks, and a complete repeat passed. The later trials were used in prior studies. This offline result does not establish a benefit in a live interface or under real stimulation.

stereoelectroencephalography · SEEG · median power · fault robustness · task-state decoding · digital fault · signal processing

Admitted 2026-11-09 19:05 UTC6 citing publications

References

usesA causal filter gave higher fixed decoder accuracy than fourfold sample removal alone in eight SEEG recordings sci:sha256:e48923d80e223e23c65d73b7a5e6734993070247d942334dabdcfd1ffa26c7b6
This upstream study supplied the source derivation and pinned EDF byte range records for the native input windows. Those records support the source audit and rebuild recipe.
usesTen spaced gaps gave lower mean zero-fill SEEG accuracy than a circular gap block at 20 percent simulated sample loss in eight fixed decoders sci:sha256:35c3a178c2212bd9b314433661dfc484678cb20f0c291b7c94bc044397f7c9b9
This study uses the exact 500-sample integer bipolar inputs, fixed early/later split, selection files, source records, and frozen full-window models. The copied files match the parent manifest hashes.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This upstream study supplied the task-state definitions, trial exclusion rules, early-trial pair selection, and fixed native decoder. The present work holds those choices fixed.
doi:10.18112/openneuro.ds004473.v1.0.2
doi:10.1088/1741-2552%2Facae0a

Cited by

6 publications

Public catalog entries on Sciscape.

usesEight fixed median gamma SEEG readers had 0.500 mean worst contact Brier error under five simulated 1 mV pulses across five segments at 10 percent active time
The result for one brief biphasic contact fault guided the pulse shape and the question about repeated faults. The present test uses fixed gamma readers from the direct parent. It does not repeat the earlier three-band comparison.
usesTwo 250-sample spectra gave lower mean Brier error than one 500-sample spectrum in eight reused SEEG records with equal sample weights
Its prior block-power fault test informed the design and scope. The present test uses unmodified inputs and a different primary comparison.
usesMedian 70–110 Hz power gave higher mean Brier error than block mean power in eight SEEG records with one simulated three-sample contact pulse at ten times early RMS
This related study supplies context for median block power, the intact-data cost, and conditional rank bounds. It used three bands and ten-sample 1 mV contact pulses. It was read after the present primary calculations and did not change the plan or fits.
usesEight offline SEEG readers with median gamma power had 0.500 mean worst contact Brier error under simulated 1 mV, 100 Hz faults of 250 samples
The prior brief-fault median result and local power bounds define the question about sustained gamma faults. This work tests longer sine faults and worst contact probability error.
usesPower blocks with 50 percent overlap increased correct SEEG outputs with a guarantee for a fault of ten samples on one contact in eight recordings at C = 1
This study uses the five-block median power method, its one-contact fault bound, its exact selected inputs, and its fixed trial and pair selection. The five-block primary result was reproduced before the overlap comparison.
usesWith separate fits, removing the strongest 100-sample block gave higher mean SEEG task state accuracy than position controls in eight patients
This study uses the exact 500-sample integer SEEG inputs, inherited pair and trial selection, source records, and block power and logistic fit code. It tests removal of recorded strongest block powers with separate reader fits and equal length position controls.