Twenty trials matched the 60-trial median AUC, but non-inferiority was shown only from 30 trials in eight SEEG subjects

We tested calibration size and within-session drift for an offline cued key-press detector in eight SEEG subjects, using archived features from OpenNeuro ds004473 v1.0.2. The fixed plan trained on the first 4–60 trials and tested on trials 61–100. The primary decoder selected eight channels on training data and used shrinkage LDA on 70–150 Hz features. Median AUC was 0.833 after 20 trials and 0.821 after 60; the median difference AUC(20) − AUC(60) was +0.003. Non-inferiority at a 0.02 AUC margin was not shown at 20 trials (one-sided p = 0.086), but was shown at 30 trials (p = 0.039) and 40 trials (p = 0.016). The all-channel decoder reached its per-subject threshold later (median 35 trials) than the eight-channel decoder (median 20). Recent calibration trials were not better than old trials, and time-ordered splits were not worse than random splits. Label-free re-centering slightly reduced AUC in the reported contrasts. These secondary tests were not corrected for multiple comparisons. This is offline evidence from one 11–19 minute session per subject; cue- and response-aligned windows, eight subjects, low power, and two near-chance subjects limit interpretation. The result does not establish performance for online onset detection or drift across hours or days.

SEEG · brain-computer interface · calibration · within-session drift · shrinkage LDA · ROC AUC · key-press decoding · OpenNeuro ds004473