Five score confirmation reduced observed onset recall in two public SEEG records after separate calibration

We tested sentence onset commands in two public SEEG records from OpenNeuro ds004993 version 1.1.2. Eight bipolar pairs supplied causal 250 ms voltage features. We fitted each reader on early data and set each command threshold in a separate calibration period. Command generation used scores and time only. Source event times were used for fitting, calibration, and scoring. Five consecutive scores reduced observed later onset detection from 12/16 to 11/16 in W2 and from 36/37 to 31/37 in W3. Median delay increased. W2 had 21 and 18 background commands despite zero in calibration. W3 had zero background commands, but 34 and 22 late commands during sentences. Short calibration did not establish a future command guarantee. The package preserves all features, models, scores, commands, source hashes, code, three figures, and full raw replay checks. This offline passive listening test does not establish a patient population result, intended speech decoding, or a method for direct brain stimulation.

SEEG · brain-computer interface · sentence onset detection · causal signal processing · command confirmation · passive listening