Delay SEEG power gave an uncertain 0.4% mean error reduction for later trials in 13 patients

We tested whether delay-period beta and high gamma power improved recall estimates for later trials in the same session. We used OpenNeuro ds006136, version 1.0.1. The data had 13 patients, 14 sessions, and 143 bipolar SEEG channels. We attached original trial IDs and used all released channels. A ridge model used trial position plus fixation and encoding power. A second model added delay power. Each channel model used earlier trials for fitting and later trials for testing. We averaged losses within patients and gave each patient the same weight. At the fixed primary penalty of 0.1, delay power reduced mean squared error from 0.415660 to 0.414011 recalled objects squared. The mean reduction was 0.001649, or 0.397%. The 95% patient bootstrap interval was -0.005700 to 0.010330. Five of 13 patients improved. The evidence does not establish a gain across patients. A model that used the past recall mean had lower error, at 0.372558. All full-sample intervals included zero at three tested penalties. The test is limited to mean band power, linear models, and source trials retained after artifact checks. The package preserves 19,617 channel trial feature rows, 2134 unique trial labels, test predictions, signal excerpts, source metadata, raw-file hashes and versions, code, and reproduction instructions.

working memory · SEEG · beta power · high gamma · recall prediction · ridge regression · forward testing · patient bootstrap