A local power reference lowered recall forecast loss compared with absolute power in only three of eight later SEEG sessions at C=0.1

This study tests a local power reference for SEEG memory readers. It uses OpenNeuro ds004809, version 2.2.0. Eight people each supply 180 later-session test words. Relative power is post-word log power minus pre-word log power. The main logistic reader uses C=0.1 and a separate target intercept adjustment. Mean feature shift fell in all eight people. Mean forecast loss fell from 0.8733 to 0.8393 nats per word compared with absolute power. However, only three people had lower loss. The mean gain was 0.0339, with a 95% person bootstrap interval from -0.0207 to 0.0939. Relative power had higher loss than task inputs in all eight people. A stronger weight penalty narrowed the excess loss. The pre-word interval contains ongoing activity and often part of the previous word display. This extension is exploratory and reuses an existing cohort. It does not test stimulation benefit. The package preserves features, source pins, code, readers, results, four figures, and reproduction instructions. Independent voltage and reader checks are included.

stereoelectroencephalography · memory recall · power spectrum · relative power · log loss · later-session prediction