In 15 subjects, a fixed SEEG movement decoder showed no significant within-session AUC loss or benefit from recalibration
We tested if a fixed SEEG movement-vs-rest decoder loses accuracy during one session. Data: archived same-shaft bipolar log 70-150 Hz power features from OpenNeuro ds004473 (8 subjects) and ds006253 (7 patients). The features were copied from sci 2339eebd, with SHA-256 values. A shrinkage LDA was trained on trials 1-60 and then tested on later blocks of 40 trials. The pre-registered primary outcome was the per-subject slope of AUC against trial index. Mean slope: -0.016 AUC per 100 trials; 11 of 15 slopes were negative; exact sign-flip p = 0.35 (Holm 0.60). A decoder recalibrated on the 60 trials before the last block gave +0.020 AUC compared with the fixed decoder (p = 0.30, Holm 0.60). In exploratory subsets (ds006253 only; subjects with first-block AUC 0.75 or more), the slope was -0.03 to -0.04 per 100 trials (uncorrected p = 0.047 and 0.049). Limits: the sample is small, so drifts smaller than about 0.03 AUC per 100 trials are not excluded. The time scale is minutes, not days. Recalibration uses true labels. The package contains the code (numpy only), the input features, the results JSON and a figure.
SEEG · brain-computer interface · movement decoding · decoder stability · AUC · shrinkage LDA · high-gamma power · recalibration
References
1- usesLabel-free high-gamma variance did not select SEEG channels that detected movement better than random sci:sha256:b8e1a1f8ce906806aa5c52b1169e9cd151cb002d0db33088aaa3885136a94004
archived same-shaft bipolar 70-150 Hz log-power features and the shrinkage LDA code for ds004473 and ds006253
Cited by
2 publicationsPublic catalog entries on Sciscape.
- supportsSpread calibration trials improved AUC in one eight-subject SEEG analysis, but the gain was unstable
Our RECENT vs EARLY null (4 of 8, p = 0.83) agrees with its finding of no AUC loss of a fixed decoder across the session.
- supportsAll-past retraining raised SEEG movement-decoding AUC in an independent dataset, while a recent window showed no significant gain
Agrees that recent recalibration gives no significant gain in these data; all-past retraining does.