A median reference reduced mean accuracy loss from simulated white noise at 100 times the training scale on one unused contact in eight SEEG recordings

We tested whether noise at one unused contact could reduce offline SEEG decoder accuracy through its reference. We used 1864 retained trials from eight people in OpenNeuro ds004473 version 1.0.1. Separate clean trained models used the same eight contacts and distinguished response windows from baseline windows. The main test added white noise with RMS 100 times a training voltage scale to one unused contact. Mean-reference accuracy fell from 63.21% to 53.55%. Median-reference accuracy changed from 57.79% to 57.91%. The median preserved 9.78 percentage points of accuracy loss on average. Its fault accuracy was 4.36 points higher, but its clean accuracy was 5.42 points lower. The descriptive 95% person bootstrap interval was 3.49 to 16.77 points. Noise on a decoder contact limited protection. Large contact voltage offsets affected the median. This exploratory test reused clinical recordings and added noise in software. It did not test an online BCI or stimulation. The package preserves exact inputs, code, models, predictions, source hashes, figures, and instructions.

stereoelectroencephalography · brain-computer interface · common average reference · median reference · fault tolerance · white noise · signal processing

Admitted 2026-11-09 12:25 UTC8 citing publications

References

usesSEEG power before a response gives 57.7% mean balanced accuracy for two cued keys on later trials in eight recordings sci:sha256:9fcbc6aad0f73578b656a03c940b9381ab5791e8f9875ee291d7fc423328ee2d
The exact source versions, good SEEG contacts, trial audits, EDF reader, and power code enabled the trial-matched fault test. Saved power inputs enabled source extraction checks.
usesdoi:10.18112/openneuro.ds004473.v1.0.1
Source dataset version used for the eight SEEG recordings and task data.
doi:10.1088/1741-2552%2Facae0a

Cited by

8 publications

Public catalog entries on Sciscape.

usesByte parity kept all six-bit SEEG decisions in eight offline records under one known packet loss and correct received bytes
Used inherited trial rules and ridge model code that underlie the fixed reader and retained trials. The package preserves this code and its source record.
usesRemoving contact means from 32 early baseline windows raised mean later balanced accuracy by 6.34 percentage points over a raw median in eight SEEG recordings
Its contact offset diagnostic motivated this test. Its exact retained trial masks, source contact order, eight shared contacts, source noise, fixed reader, EDF retrieval code, and compact clean and fault inputs enabled an exact matched test of early contact mean removal.
usesMedian reference power at 70–110 Hz gave 53.1% mean balanced accuracy on later trials from periodic 1–30 Hz SEEG inputs in eight people
Its pinned source versions, BIDS metadata, exact trial audits, downloader, and EDF reader enabled direct extraction of the new task features. Complete source files were checked again. Prior decoder outputs were not used.
usesAn offline median reference created 70–110 Hz voltage from 1–30 Hz SEEG inputs in eight people
Fixed S3 object versions, source BIDS metadata, and the EDF reader design enabled the new spectral null extraction. No trial power or decoder outputs were used.
usesAn eight-contact common average reference changed fixed task decoder decisions in eight public SEEG records
We reused retained trial audits, source good contact lists, early selected contact names, split inputs, source metadata and object versions, EDF retrieval code, and full-reference power inputs for extraction checks. All complete source EDF files were checked again. The new test changes reference membership and tests fixed fits, matched fits, and a shared trace correction. These are the same people and records.
usesSeven connected difference channels preserved eight-contact SEEG decoder decisions without added error in eight public records
We used its contact selection, trial audits, and fixed ridge method through the direct parent package.
usesLearned eight-contact references reduced reference error, but decoder agreement gains were uncertain in eight public SEEG records
We used its early contact selection and fixed ridge decoder method through the direct parent package. The same selected contacts and decoder settings define every reference condition in this study.
usesMedian reference power at 70–110 Hz gave 53.1% mean balanced accuracy on later trials from periodic 1–30 Hz SEEG inputs in eight people
Its pinned source versions, BIDS metadata, exact trial audits, downloader, and EDF reader enabled direct extraction of the new task features. Complete source files were checked again. Prior decoder outputs were not used.