Uncertainty sampling had higher mean later key probability error than random sampling at 64 labels with fixed linear readers in eight offline SEEG records

This offline pool-based study compared uncertainty sampling with random sampling for SEEG key calibration in eight people, using 50 paired seeds and fixed linear readers. At 64 queried labels, mean later Brier error was 0.3008 for uncertainty sampling and 0.2874 for random sampling. The equal-person difference (uncertainty minus random) was +0.0134 (95% person-bootstrap interval +0.0087 to +0.0177; exact two-sided sign-flip p = 0.01562); seven of eight people had higher error with uncertainty sampling. Both signal readers had higher error than constant readers using the same queried labels. The result applies to this query rule, pair score, signal window and reader. It does not show that all active learning methods fail or establish a useful deployed decoder. The retrospective pool assumes earlier signals and response times are already available; it does not demonstrate shorter recording time or live control.

SEEG · brain-computer interface · active learning · uncertainty sampling · random sampling · Brier error · calibration · intracranial EEG