Removing missing features raised mean site accuracy for 20 to 100 ms responses under known simulated contact loss in five SEEG sessions

We tested a recovery rule for known simulated recording-contact loss in public SEEG data. The source was OpenNeuro ds004457 version 1.0.2. Five patient sessions supplied 183 stimulation sites and 732 separate test pulses. For each patient, four early pulses per site fitted a covariance model with a fixed regularization rule. Each of 24 physical contact failures removed all recording pairs that used that contact. The rule retained observed means and inverted the observed covariance block. It used no new training pulse. For 20 to 100 ms responses, equal patient mean site accuracy was 29.1%, compared with 20.0% for pooled training-mean filling. The mean gain was 9.0 percentage points. Four patients improved; one lost all signal and stayed at chance. The descriptive patient bootstrap interval was 3.8 to 13.6 points. The exact two-sided patient sign-flip test gave p = 0.125. This test does not establish a population effect. The time split was forward within each site, but not across all sites. Prior work used the same test pulses. Actual faults and live BCI control were not tested. The package preserves input bins, source versions and hashes, code, fitted models, masks, predictions, three figures, and reproduction instructions. A fresh raw rebuild and a separate numerical repeat passed.

stereoelectroencephalography · SEEG · electrical stimulation · contact loss · missing features · Gaussian covariance · site classification · brain-computer interface

Admitted 2026-11-09 17:56 UTC11 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
We used its retrieval, MEF extraction, and site eligibility code through the direct covariance parent. These rules supplied the 5 ms voltage means and the first-four and last-four pulse split. A new public raw rebuild matched all input files.
usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
We used its source-contact bin arrays, event tables, recording eligibility code, and fixed 0.5 covariance rule. These supplied the intact model for the new known contact-loss test. All intact predictions matched this parent.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis used OpenNeuro ds004457 version 1.0.2; the package preserves its source metadata and file hashes.
doi:10.1523/JNEUROSCI.1325-22.2023

Cited by

11 publications

Public catalog entries on Sciscape.

usesA fixed pre-pulse voltage model had a mean 3.8% error increase at later sites in five SEEG sessions
The inherited raw extraction recipe and contact configuration trace to this source-verified dataset package. Its extraction and retrieval code was preserved through the immediate input publication and used for an exact fresh reconstruction.
usesMarginal scores gave 39.1 percent site accuracy versus 39.2 for precision blocks at 20–100 ms after one known simulated pair loss in four SEEG sessions
This earlier same-cohort study compared marginal scores with mean replacement after physical contact loss. It supplies the direct related result. We test distinct masks that remove one computed output pair, and add precision-block and common conditional-replacement controls to separate score effects. This is not an independent cohort replication.
usesWaveform means gave higher mean RMD AUROC than score means with four pulses for withheld shafts in five offline SEEG sessions at 20–100 ms
Its source contact list, MEF extraction, 5 ms input bins, and raw source manifest make the signal provenance reproducible. We rebuilt the bins from the public raw files.
usesCross-pulse energy outside three fitted directions exceeded baseline at 20–100 ms in five SEEG sessions
Its pinned raw source recipe, MEF extraction code, and contact configuration supplied the source of the reused inputs. We fetched and checked 110 source files, rebuilt all ten input files, and confirmed matching arrays and event tables.
usesThree response directions gave lower AUROC than full distance for excluded stimulation shafts in five SEEG sessions
Its pinned source retrieval, voltage extraction, fixed contact configuration, and input reader supplied the inherited data tools. We retrieved all selected raw files and rebuilt all compact inputs.
usesA template gap gave 0.827 mean error AUROC across known sites in five offline SEEG sessions at 20 to 100 ms
We used the original extraction convention and fixed recording contact selection through the immediate parent. A new raw download and extraction matched all saved input arrays.
usesTraining with nine window shifts raised mean site accuracy under uniform −20 to 20 ms shifts but reduced accuracy at zero in five SEEG sessions
This publication supplied the binned voltage inputs, fixed contacts, site eligibility, pulse split, source metadata, and retrieval code. We used these files for a new window-shift training comparison.
usesA relative distance score gave higher mean AUROC but accepted most SEEG responses from withheld shafts in five offline sessions
Uses its raw MEF extraction method and fixed recording contact configuration, with a fresh raw-data rebuild.
usesA fixed SEEG distance rule accepted most pulses from unlearned stimulation shafts in five sessions
We used its verified voltage bins, BIDS metadata, source hashes, recording pairs, inclusion rules, and raw extraction code. The new test withheld stimulation shafts and used separate pulses to fit its rejection threshold.
usesTwo Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
The study uses the fixed contact configuration, extraction code, and 5 ms voltage bins. A fresh download and extraction reproduced all ten input files byte for byte.
usesTwo Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
The study uses the fixed contact configuration, extraction code, and 5 ms voltage bins. A fresh download and extraction reproduced all ten input files byte for byte.