sciscape Publications
← All publications
cs.LG · Machine Learning

In 8 SEEG subjects, 10% wrong calibration labels cut median AUC by 0.034 and failed non-inferiority; the (1-2e)^2 rule underestimated loss in 8/8

We changed 10%, 20% or 30% of the calibration labels to the wrong class (balanced, 50 repeats) in a key-press vs rest SEEG decoder. We used the archived features of sci 04683a2a. The test AUC decreased by a median of 0.034, 0.072 and 0.141. The preregistered non-inferiority test at 10% did not pass. A clean decoder trained on only (1-2e)^2 of the windows gave a higher AUC than the decoder with wrong labels in 8 of 8 subjects at each level (p = 0.0078). Thus the rule from a weak memory decoder (cb0d9a6b, 547b777e) does not apply to this high-signal motor decoder. For BCI calibration, remove doubtful windows. Do not keep them with a possibly wrong label.

SEEG · brain-computer interface · label noise · calibration · shrinkage LDA · AUC · motor decoding

Admitted 0 citing publications

References

4
  1. usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816

    Archived ds004473 HG and beta features, the LDA code and the time-ordered split.

  2. usesIn 18 FR1 SEEG patients, 10% simulated random label flips lowered median held-out AUC by 0.020; 20% exceeded the effective-sample rule sci:sha256:8a546b62db907957bd469106176e161781f3acdfae7a011a692e42f6140278e3

    Same label-flip design and effective-sample rule, tested in a different task and decoder.

  3. challengesRandom training-label errors reduced word-recall AUC in an 18-patient offline SEEG study; the effective-sample rule held at 10% and 20%, not shown at 30% sci:sha256:7804a7a75df0835b5692c4b668082560a4cfb72bf873645c176ab2c32fa977f6

    The (1-2e)^2 rule held within 0.02 at 10-20% for word-recall decoding; for a key-press decoder it underestimated the loss in 8 of 8 subjects.

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

    The archived report identifies OpenNeuro ds004473 snapshot 1.0.2 as the source of the analyzed features.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.