In 8 SEEG subjects, calibration CV did not significantly overstate held-out ITR; oracle thresholds and window-only timing raised estimates

We compared three choices that can raise information transfer rate (ITR) estimates for an offline, cued SEEG key-press versus rest switch: calibration-set cross-validation (CV), an oracle threshold selected on scored data, and analysis-window time instead of task time. Archived features from 8 subjects (OpenNeuro ds004473 v1.0.2) were classified with shrinkage LDA. The preregistered primary contrast was the CV oracle estimate (trials 1–100, 5-fold CV) minus held-out ITR (train on trials 1–100, test on trials 101+, fixed threshold), both using a 0.5 s decision time. The mean difference was +7.1 bits/min (median +0.8); it was positive in 4 of 8 subjects and the exact one-sided sign-flip p-value was 0.19. The hypothesis that CV oracle ITR overstates held-out ITR was not supported at group level, although subject differences reached 28–41 bits/min in two subjects. In descriptive secondary analyses, oracle thresholds raised ITR in all 8 subjects (median +2.1 bits/min in CV and +4.9 in test). Using task time instead of the 0.5 s window reduced median held-out ITR from 28.7 to 3.5 bits/min (ratio 7.7; median task time 3.83 s per decision). The analysis was offline, used one session per subject (n=8), balanced classes, fixed 0.5 s windows, and a Wolpaw formula that assumes equal error rates and no error correction; task-time estimates reflect this experiment’s rhythm, not a real BCI.

brain-computer interface · information transfer rate · SEEG · intracranial EEG · cross-validation · linear discriminant analysis · threshold selection · held-out evaluation · Wolpaw ITR