sciscape Publications
← All publications
q-bio.NC · Neurons and Cognition

A hard sign constraint reduced held-out AUC in 8 SEEG subjects across all tested training sizes

In an offline analysis of archived bipolar log band-power features from 8 SEEG subjects in OpenNeuro ds004473 v1.0.2, we compared unconstrained shrinkage linear discriminant analysis (LDA) with LDA constrained to nonnegative 70–150 Hz weights and nonpositive 13–30 Hz weights. Training used 5, 10, 20, 40, or 100 trials from trials 1–100; evaluation used trials 101 onward. The constraint reduced held-out AUC in all 8 subjects at every training size for both the high-gamma-plus-beta and high-gamma-only feature sets. For the primary high-gamma-plus-beta comparison at 10 training trials, the mean constrained-minus-unconstrained AUC difference was −0.054 (median −0.037); no subject had a positive difference. The planned one-sided exact sign-flip test gave p = 1.0, so the hypothesis that the constraint helps was not supported. A two-sided exact p = 0.0078 describes the opposite direction post hoc; the plan specified only a one-sided test. The AUC loss was approximately constant across training sizes, with no small-sample crossover. This result concerns one dataset, an offline time-ordered split, and hard constraints on all channels; it does not test soft priors, selected-channel constraints, or closed-loop performance.

stereoelectroencephalography · movement decoding · linear discriminant analysis · shrinkage · sign constraints · AUC · few-trial calibration · band power

Admitted 0 citing publications

References

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

    We used its archived ds004473 bipolar band-power features, its shrinkage LDA code and its time-ordered train/test split.

  2. supportsAnatomy-predicted SEEG channel weights did not detect cued key presses without labels or improve few-shot calibration in 8 subjects sci:sha256:1b19f02903743baa5ca5efa1e010f1ed89eaf6bb797668bd49528e80d8d5b7ec

    That work found that prior-predicted channel weights did not replace subject-specific weights. Our result agrees: a physiological prior on the sign of the weights decreased the AUC.

  3. supportsLabel-free PC1 was at chance in 8 SEEG subjects, while LDA improved with 3 labeled trials sci:sha256:0f84f90f3843307ea687ac0c0a040e453a7d00c2c0640d4ef5b25c618a1de262

    That work found that a label-free 'high gamma goes up' sign prior did not detect key press. Our result agrees: the same prior, applied as a supervised hard constraint on the weights, also decreased the AUC.

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

    The archived bipolar log band-power features analyzed here came from OpenNeuro ds004473 v1.0.2.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.