Adding forecast angle agreement lowered mean Brier score for 50 ms offline theta phase labels in seven SEEG rest records

We tested whether two phase forecasts can give a better confidence estimate when their angles agree. The data were seven public SEEG rest records from OpenNeuro ds006065 version 1.0.0. We used four selected contact pairs per person. Two readers used 100 ms and 500 ms of past voltage. They forecast a phase label 50 ms after each cutoff. Early data fit the readers and confidence models. The later test had 96,320 forecast opportunities. Adding cosine angle agreement to both log output lengths lowered mean Brier score from 0.23024 to 0.22916. The mean reduction was 0.00108. The descriptive 95 percent person interval was 0.00042 to 0.00176. Six of seven people had lower scores. The exact two-sided sign-flip value was 0.03125 under independent people and symmetric null differences. A post-primary quadratic strength check also gave a reduction. The strictest gate slightly lowered successful yield. The package includes selected voltages, all forecasts, model weights, three figures, a reader, and reproduction code. Fresh full downloads matched all seven source hashes and every saved voltage. The cohort and later periods were reused. The phase label uses future samples. This study does not test delivered pulses, hardware timing, or brain control benefit.

stereoelectroencephalography · theta phase · forecast confidence · angle agreement · Brier score · phase reader