A 0.1 s SEEG band-power window lowers AUC versus 0.5 s but raises the bit-rate upper bound in an offline key-press task

We measured the trade-off between analysis window length and decoding performance for cued key-press detection in stereo-electroencephalography (SEEG). The analysis used the first 100 trials from each of 8 subjects in OpenNeuro ds004473 v1.0.2, good contacts in an adjacent bipolar montage, log band power at 70–150 Hz (high gamma, hg) and 13–30 Hz (beta), shrinkage LDA, and 5-fold cross-validation. Filtering used a fixed 1.5 s segment and an offline, non-causal zero-phase filter; window lengths of 0.05–0.5 s refer to power integration, not causal latency. In the pre-registered hg test, AUC(0.1 s) minus AUC(0.5 s) averaged −0.063 (median −0.073), and 7 of 8 subjects lost AUC. Non-inferiority at margin 0.05 was not shown (seed-0 p = 0.74). Across 10 CV seeds, the mean difference was −0.067 (p = 0.86); the 0.5 s window was better in all 8 subjects (p = 0.004). The 0.1 s window produced a higher hg Wolpaw bit-rate upper bound than the 0.5 s window in 7 of 8 subjects (median increase 0.82 bits/s, p = 0.008). Windows of 0.3 s were non-inferior to 0.5 s for all three feature sets (p ≤ 0.008). These results apply to this offline, balanced, two-class task. The move window was centred at key-press onset and therefore includes post-onset data; this is not a movement-prediction result. The bit-rate upper bound assumes one decision per window, no gaps, and does not account for false alarms. The sample includes 8 subjects, one task and clinical electrode placements.

stereo-electroencephalography · SEEG · brain-computer interface · key-press detection · band power · window length · AUC · bit rate · shrinkage LDA · non-inferiority