Within-session transfer AUC was similar in both time directions in eight SEEG subjects

This analysis asked whether a key-press-versus-rest decoder transfers differently when trained on early trials and tested on later trials than in the reverse direction. It used archived features from eight subjects (132–298 trials per subject) in OpenNeuro ds004473 version 1.0.2. Features were log mean power in high gamma (70–150 Hz) and beta (13–30 Hz) bands on 100–112 adjacent-bipolar SEEG pairs. Each session was split into its first and second halves. Shrinkage linear discriminant analysis (LDA) was trained on one half and evaluated by area under the ROC curve (AUC) on the other. The primary endpoint was backward minus forward AUC with high-gamma and beta features; an exact two-sided sign-flip test used all 256 patterns across subjects. The mean difference was +0.003 AUC (median −0.002; 3 of 8 subjects above zero; p = 0.75). With high-gamma features alone, the mean difference was −0.004 (median +0.002; 5 of 8 above zero; p = 0.73). The largest absolute difference was 0.046 with both bands and 0.061 with high gamma alone. The within-half five-fold cross-validation AUC for the second half was also not consistently different from the first (both bands: mean change −0.039, 4 of 8 higher, p = 0.19). These data do not show a direction effect, but this non-significant result is not evidence of equality. The analysis included only eight subjects, one session per subject, and offline decoding; an effect of about 0.02 AUC cannot be excluded. The halves can differ in trial count, and their time separation was not controlled.

stereoelectroencephalography · brain-computer interface · decoder transfer · shrinkage linear discriminant analysis · ROC AUC · high-gamma power · within-session analysis