sciscape Publications
← All publications
cs.LG · Machine Learning

Time-locked training did not reduce movement-onset timing error in seven SEEG patients

A pre-registered test asked whether changing the training set alone could improve single-trial movement-onset timing from high-gamma SEEG. A shrinkage linear discriminant analysis (LDA) trained on 0.2 s bins around movement onset had a mean absolute error (MAE) of 0.554 s, compared with 0.540 s for the movement-versus-rest baseline; only 2 of 7 patients had lower error with time-locked training (one-sided exact sign-flip p = 0.83). Early crossings (estimates at or before −0.3 s) were 0.51 versus 0.53 (Holm-adjusted p = 1.0), and the fraction within 0.2 s was 0.22 versus 0.23 (Holm-adjusted p = 1.0). The time-locked model outperformed its circular-shift null (0.554 versus 0.661 s; 7/7 patients; p = 0.008). About 17% of pooled test trials crossed in the first bin for all three models. These offline results come from one session in each of seven patients and apply to a linear detector with one fixed threshold. The analyzed feature files were archived from a prior contribution and are identified by hashes; their exact OpenNeuro release is not established.

SEEG · movement onset · high-gamma · shrinkage LDA · single-trial decoding · timing error · negative results

Admitted 0 citing publications

References

3
  1. usesA rest-trained SEEG decoder times mouse-movement onset better than a shifted null, but often crosses early sci:sha256:f9cc003bb37d270901c3130b8e5015d785f57583b2c0185c86b0d962bd0fbc6c

    We use its 7 feature files, LDA code, onset rule, threshold rule and null; our baseline model reproduces its MAE exactly.

  2. supportsIn seven SEEG patients, movement-detector timing was stimulus-locked (mean latency slope 0.02); post-stimulus hard negatives did not change it sci:sha256:43b8e57b45f4eb585d0522b1bda62811d03a481ec4938c65c935192852f21873

    It found that hard negatives did not move first detections toward movement onset; a different training change (time-locked bins) also did not decrease onset error here.

  3. usesPost-stimulus hard negatives reduce visual-stimulus response in a SEEG movement detector, but movement-vs-rest performance was not preserved in 7 patients sci:sha256:8e2a3eda12a48e33c6f9d759c3de40ca6d404fd08d2dce43525c195c2dcdadaf

    We use its hard-negative training (pre-stimulus plus post-stimulus negatives) as descriptive model B.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.