Spread calibration trials improved AUC in one eight-subject SEEG analysis, but the gain was unstable

We compared three ways to choose N labeled calibration trials for an offline SEEG key-press detector: the first N trials (EARLY), N trials spread across the calibration block (SPREAD), or the last N trials before testing (RECENT). We used archived high-gamma (70–150 Hz) and beta (13–30 Hz) features from eight subjects in OpenNeuro ds004473 v1.0.2 and shrinkage LDA. The held-out test block was the last third of each subject’s session. The analysis plan was written and hashed before analysis. For the planned primary contrast at N = 40, SPREAD exceeded EARLY by a mean 0.032 AUC (7 of 8 subjects; exact two-sided sign-flip p = 0.023, significant after Holm correction). RECENT exceeded EARLY by 0.011 AUC (4 of 8; p = 0.83). The SPREAD result was not stable: it was not significant at N = 10, 20, or 60, or with high-gamma features only. In an exploratory analysis, shifting the evenly spaced grid reduced the mean gain to 0.017 AUC and the median to 0.001 (4 of 8; p = 0.38). One subject contributed most of the planned-grid gain: its first 40 trials yielded AUC 0.62, versus 0.73–0.83 for the other rules. These data do not support a general rule to spread calibration trials or to choose trials near the time of use. They show that calibration from only the first trials can produce a poor decoder for some subjects. The evidence is limited to an offline analysis of one session each from eight epilepsy patients in one dataset; secondary and exploratory tests were not corrected for multiple comparisons, and the shifted-grid analysis was exploratory. Non-significant results do not establish equality.

SEEG · brain-computer interface · calibration · linear discriminant analysis · AUC · high-gamma power · beta power · trial selection