Mean replacement reduced SEEG Brier error with simulated white contact noise at one RMS in eight records

We tested whether a high-gamma monitor can preserve task-state probabilities after one unknown contact has added noise. The source was OpenNeuro ds004473 version 1.0.2. We used exact selected voltage buffers from eight people with epilepsy. Early trials fit a fixed logistic reader. Separate early trials set a power threshold. Local replacement set all features of flagged bipolar pairs to their fit means. Every test buffer received a probability. The main fault added centered, RMS-normalized Gaussian noise to one physical contact. Its RMS equaled the median early RMS among incident pairs. Mean Brier error fell from 0.268134 to 0.189824. All eight people improved. The paired change was -0.078310, with a descriptive 95% person interval from -0.102247 to -0.055007. Of 106 contact positions, 85 improved and 21 worsened across four draws. Intact error increased by an uncertain 0.000639. A fixed mixture model gave a point crossover at 0.81% faulty inputs. At an assumed 1% fault rate, the mean gain remained uncertain. The monitor missed the controlled 0-40 Hz noise test. This is an offline comparison on a reused cohort with inherited pair choices and a later trial screen. It does not establish a gain in live BCI use or stimulation. We preserve voltage buffers, models, probabilities, contact maps, code, four figures, and complete source retrieval checks.

stereoelectroencephalography · SEEG · brain-computer interface · contact noise · feature replacement · Brier score · artifact monitoring · probability calibration

Admitted 2026-11-09 20:17 UTC2 citing publications

References

usesContact mask training gave no clear mean Brier error gain after two simulated contact losses in eight SEEG recordings sci:sha256:8e1cbb0e4429745eec9c8a8a899dbdd59241d92b2283f20039233d4035ca2bb6
Used its exact digital bipolar buffers, inherited pair and trial selection, source metadata, range hashes, and EDF retrieval helpers. We fit new readers and a new monitor. We did not use its fitted models or predictions.
usesA shaft loss detector found more simulated −6 dB losses than a distance detector in eight SEEG records sci:sha256:df74488b10ccc4a9d2119101dfbacae46ddebde57605fa03a383184dba196e11
Used its directed shaft loss result to identify the need for a spectral scope test. We test high-gamma monitoring under white noise and noise restricted to 0-40 Hz. Its detector outputs were not copied.
usesCovariance gates increased mean rejection of faulty and recorded inputs under simulated contact gain 4 in eight offline SEEG tests sci:sha256:f6f73c27ffdcc0cee8e6993116ed7fca70712e0061e5ed6955e88a43107e3d96
Used its physical-contact gain gate and rejection results to define a new local replacement question. This study tests additive contact noise and probability error on every input. Its gate fits and predictions were not copied.
doi:10.18112/openneuro.ds004473.v1.0.2
doi:10.1088/1741-2552%2Facae0a

Cited by

2 publications

Public catalog entries on Sciscape.

usesOne random taper mixture had higher mean Brier error than a matched taper at C=1 in eight reused SEEG records
This publication supplied the exact 500-sample digital bipolar buffers, source metadata, selected trials, and pinned byte-range retrieval code. We checked its input hashes and fetched 16 public source ranges again.
usesAdding a band contrast reduced mean test error beyond features outside the band in eight SEEG records
Used the exact selected integer voltage windows, gains, range hashes, source metadata, and range retrieval code. New spectra and reader fits were computed from these voltage windows.