Across 8 offline SEEG subjects, random 50-trial calibration remained non-inferior on average to 100-trial calibration

We tested whether a previously reported non-inferiority result for a cued SEEG key-press detector depended on calibrating with the first 50 trials. In eight subjects from OpenNeuro ds004473, we used the archived per-trial bipolar log power features and the same NumPy shrinkage linear discriminant analysis (LDA) as the prior SCI contribution. For each subject, we evaluated 200 random subsets of 50 trials from trials 1–100 on the unchanged test set of trials 101 and later. The primary contrast was AUC for 70–150 Hz high-gamma plus 13–30 Hz beta power on a random 50-trial subset minus AUC for high-gamma power calibrated on trials 1–100; the non-inferiority margin was −0.02 AUC. The mean contrast was +0.020 (median +0.005), and all 8 subjects exceeded the margin (exact one-sided sign-flip p = 0.0039). Removing any one subject gave p = 0.0078. With 30 random trials, non-inferiority was not shown (mean +0.002, p = 0.074). Across subjects, 0.62–1.00 of random 50-trial draws were non-inferior, so a single calibration draw could fail. Trials 51–100 did not outperform trials 1–50 (mean difference +0.007, p = 0.64). Adding beta power at 50 random trials gave a mean AUC gain of +0.037 (p = 0.016, uncorrected). Results are offline, from one session in 8 subjects, using cued fixed windows that include the press and random subsets from the same 100-trial pool. The high-gamma reference was at chance for one subject. The archived features are hash-checked, but their exact OpenNeuro release is not established.

stereo-electroencephalography · SEEG · brain-computer interface · calibration · non-inferiority · high-gamma · beta power · linear discriminant analysis · area under the curve · random sampling