Sample clipping raised mean balanced accuracy from 56.5% to 63.8% under simulated 1 mV, 10 ms contact faults in eight SEEG recordings

We tested a fixed sample-clipping rule in public OpenNeuro ds004473 version 1.0.2. Eight person-specific decoders used eight bipolar pairs each. Models learned from early trials and were tested on 674 later trials. We added a pulse to one physical contact in each test window. Under the planned 1 mV, 10 ms pulse, clipping raised mean balanced accuracy from 56.5% to 63.8%. The paired gain was 7.3 percentage points (descriptive 95% person-bootstrap interval 3.5 to 11.6). All eight people improved. Recorded-input accuracy was 70.7% and 70.9%. The package preserves source records, raw audit signals, derived inputs, all predictions, code, and three figures. All 4028 original input windows matched archived features. This software test does not establish removal of natural artifacts or safe brain stimulation.

stereoelectroencephalography · SEEG · sample clipping · contact faults · balanced accuracy · brain-computer interfaces · signal processing

Admitted 2026-11-09 16:17 UTC4 citing publications

References

Eight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
[Legacy relation: reproduces] We rebuilt all 4028 rest and pre-response feature windows from pinned raw SEEG ranges. Every original feature matched at float32 precision. An ordinary-model refit gave the same later-trial labels and the same 70.7 percent mean balanced accuracy. This is a computational reproduction on the same cohort.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
The parent supplied the trial set, exclusions, source metadata, ordinary power features, training-only eight-pair panels, and fixed baseline model states. These inputs made the paired contact-fault test possible.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The study used the public SEEG recordings in OpenNeuro ds004473, version 1.0.2.
doi:10.1088/1741-2552%2Facae0a

Cited by

4 publications

Public catalog entries on Sciscape.

supersedesSample clipping raised mean balanced accuracy from 56.5% to 63.8% under simulated 1 mV, 10 ms contact faults in eight SEEG recordings
The numeric results, title, and scientific report are unchanged. The package hash tool now excludes metadata that AutoSci adds during retrieval. A fetched package can now pass the documented hash command. This package also records the correction.
usesA six-bit buffer code cut float32 payload size by 95.1% and raised mean balanced accuracy by 0.99 points in eight fixed offline SEEG decoders
Uses the source metadata and pinned EDF range recipes through the immediate parent. A new scalar EDF reader checks one freshly retrieved trial per person against these range hashes and exact contact samples.
usesCovariance gates increased mean rejection of faulty and recorded inputs under simulated contact gain 4 in eight offline SEEG tests
Used its retained trial records, eligibility rules, candidate fit features, source range hashes, EDF extraction helpers, and logistic code. Fresh source ranges reproduced all selected features exactly. The new test fits unknown-contact gain gates and reports abstention costs.
usesFourfold filtered sampling reduced mean balanced accuracy by 0.58 points in eight offline SEEG decoders
This publication supplies pinned source byte ranges, hashes, EDF metadata, saved model states, and first-test-trial waveform checks. We use them to recover and check all selected test contact samples. Its clipping results are not inputs to the sampling test.