An onset-trained SEEG decoder detected later movement in 7 patients, but non-inferiority to a late-trained decoder was not shown

In an offline analysis of seven patients from OpenNeuro ds006253, we tested whether a movement-versus-rest decoder trained on bipolar SEEG high-gamma activity (70–150 Hz) from 0–0.4 s after mouse-movement onset detected activity at 0.6–1.0 s. Trials were ordered by stimulus onset; the first 60% trained a shrinkage LDA decoder and the last 40% tested it. The onset-trained decoder detected the later window in all seven patients (mean AUC 0.749; Holm-adjusted p=0.023). A decoder trained on the later window had a mean AUC advantage of 0.025 (worst patient difference 0.092); non-inferiority of the onset-trained decoder at a 0.05 margin was not shown after Holm correction (raw p=0.0469; adjusted p=0.0938). A secondary, uncorrected analysis restricted to test trials with movement duration at least 1.0 s found a mean difference of +0.024, with all seven patients inside the margin (p=0.0078). The onset-trained decoder’s AUC changed from 0.789 at onset to 0.749 in the later window; evidence of a loss was not shown (adjusted p=0.1406). This is an offline result from one task, one time-ordered split, one linear decoder, and seven patients. The code also excluded trials with missing values in any descriptive 0.2 s bin, beyond the preregistered exclusion rule; this deviation was found after the run and not changed.

SEEG · brain-computer interface · movement decoding · high-gamma activity · shrinkage LDA · AUC · motor decoding