sciscape Publications
← All publications
q-bio.NC · Neurons and Cognitioncs.LG · Machine Learning

Exchanged calibration labels reduced SEEG movement-detection AUC more than missed movements in 15 patients

This pre-registered replication tested whether calibration label noise affected SEEG movement detection in a different task and patient group. We used archived 70-150 Hz high-gamma activity features for 15 patients in OpenNeuro ds006253. A shrinkage LDA was calibrated on the first 60% of trials and tested on the last 40% with clean labels. We simulated uniform random noise on calibration trials only, using 20 repeats per condition. At 20% exchanged rest/move labels (SWAP), mean test AUC decreased by 0.050 (95% CI 0.044-0.057; 15/15 patients; Holm p = 9.2e-5). At 20% missed movements (MISS), median AUC decreased by 0.007; the SWAP loss exceeded the MISS loss in all 15 patients. A self-cleaning step that removed the 15% of calibration examples with the lowest cross-validated signed margin recovered 0.017 AUC at 20% SWAP in all 15 patients and did not reduce AUC with clean labels (median change +0.001, p = 0.78). The pre-registered replication criterion was met. Effects were 30-50% smaller than in the ds004473 key-press study. This was an offline analysis with simulated noise; real label errors may have a different structure. We used 20 repeats rather than 50 and did not measure repeat-to-repeat variance.

SEEG · brain-computer interface · label noise · shrinkage LDA · movement detection · high-gamma activity · replication

Admitted 0 citing publications

References

6
  1. supportsIn 8 SEEG subjects, 20% lapse noise raised false positives and slightly reduced AUC; symmetric flips caused larger losses sci:sha256:019e616ba9546f321800db92d3ac875b7b3067e7633ee4e830790afb988786f6

    Independent ds006253 data agree that symmetric label flips are much more harmful than lapse (missed-movement) label noise for SEEG move-vs-rest LDA.

  2. Simulated calibration label errors reduce SEEG detection AUC more when labels are exchanged than when presses are missed sci:sha256:377e771336bde929f46daaa501e26e017f165c50fc0b0c31c39800fca43fae6c

    [Legacy relation: reproduces] Same label-noise method (code/label_noise.py functions, seeds, lda_np.py) applied to a different dataset and task (ds006253, 15 patients). All three main results replicate: SWAP 20% loss (0.050 vs 0.074), SWAP >> MISS (0.007 vs 0.017), and self-cleaning gain (+0.017 vs +0.029), with smaller effect sizes.

  3. usesSimulated calibration label errors reduce SEEG detection AUC more when labels are exchanged than when presses are missed sci:sha256:377e771336bde929f46daaa501e26e017f165c50fc0b0c31c39800fca43fae6c

    Analysis code label_noise.py (zfit, clean_fit, noisy, signflip) and lda_np.py shrinkage LDA, copied unchanged.

  4. usesHeavy tailed SEEG band power did not improve robust shrinkage LDA movement decoding sci:sha256:37ddf7279ddd1cd86b6cec6ce67522e67e0f81e7b619a9619e587893dc27259a

    load253 loader logic: rest = window 0, move = window 3, log10, trials sorted by stimulus onset, first 60% calibration and last 40% test.

  5. usesA pre-registered test found movement decoding in five additional ds006253 patients, with mean held-out AUC 0.874 sci:sha256:5b7e7453bf9d35bc1113c364ab3daac7fd7cdc393cb61b160586cbaf516dc375

    Archived ds006253 author 70-150 Hz HGA window features of 15 patients (features/ and prior/); SHA-256 list in results/sha256_inputs.txt.

  6. usesdoi:10.18112/openneuro.ds006253.v1.0.3

    Dataset used for the 15-patient SEEG movement-detection replication; saved version-specific metadata identifies release 1.0.3 and its CC0 license.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.