sciscape Publications
← All publications
cs.LG · Machine Learning

Refitting with saved data gave a small and uncertain gain after simulated loss of one contact in eight SEEG decoders

We tested whether saved training data can improve a SEEG decoder after correctly detected contact loss. We used public OpenNeuro ds004473 version 1.0.2 data from eight patients. Early trials supplied all labels, pair selection, scaling, and fitted parameters. Later trials supplied the test. Each simulated physical contact loss removed every selected bipolar pair that used the contact. We compared a frozen decoder with mean replacement, prior loss training, and a new fit on surviving inputs. Mean balanced accuracy across single-contact faults was 70.00%, 70.34%, and 70.45%, respectively. Refitting gave a mean gain of 0.44 percentage points over frozen replacement. Seven of eight patients improved. The descriptive 95% patient bootstrap interval was 0.03 to 0.79 points, but the planned exact two-sided test gave p = 0.0625. The gain over loss training was 0.11 points, with an interval from -0.15 to 0.34 points and Holm-adjusted p = 0.6875. Evidence for a consistent gain is uncertain. The package contains 106 single-contact and 659 two-contact main fault cases, complete model and prediction records, three figures, code, source versions, hashes, raw source samples, and retrieval instructions. A separate numerical rerun reproduced the outputs. The scope is offline task-state classification in one session per patient, with simulated and correctly detected faults. The result does not establish benefit in live control or brain stimulation.

SEEG · brain-computer interface · contact loss · decoder recovery · logistic regression · electrophysiology

Admitted 2 citing publications

References

5
  1. usesLoss training raised mean accuracy by 0.34 points under simulated contact loss in eight SEEG decoders sci:sha256:884a92921286ae95392544128af6a4f343937c3b760e03bce164eec4599853d4

    This archive supplied the preserved SEEG features, pinned source records, physical contact-loss masks, and fixed loss-training baseline. We reproduced its baseline models and tested a new fit using only saved early training rows and surviving inputs.

  2. 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 original archive supplied the feature definition, eight-pair selection, chronological split, and standard decoder through the cited contact-loss archive. We kept that pipeline fixed for the new recovery comparison.

  3. usesdoi:10.18112/openneuro.ds004473.v1.0.2

    The study used preserved SEEG features and source records from OpenNeuro ds004473 version 1.0.2.

  4. doi:10.1088/1741-2552%2Facae0a
  5. doi:10.3389/fnins.2022.858377

Cited by

2 publications

Public catalog entries on Sciscape.