Adding high gamma lowered loss for separate pre-key and pre-fixation SEEG windows in fixed tests of later trials from eight patients

We tested whether high gamma power improves classification of one SEEG window without a matched test comparison. We used 2,030 aligned trials from eight patients in public OpenNeuro ds004473, version 1.0.1. Each trial supplied 500 samples before the key press and 500 samples before fixation. The first half of retained trials trained each patient model. Later trials tested it. The logistic models used frozen training constants, a free intercept, and coefficient penalty 1. Adding 70-110 Hz power to 13-30 Hz power lowered test loss in all eight patients. The equal-patient mean loss fell from 0.6238 to 0.5724 nats per window. Mean gain was 0.0514 nats, with a descriptive 95% patient bootstrap interval of 0.0263 to 0.0855. The exact two-sided sign flip p value was 0.0078125 under independent patients and symmetric null gains. Mean balanced accuracy increased from 66.11% to 70.92%. All six fixed sensitivity checks had positive mean gains. A weaker penalty gave negative gains in two patients. Full pinned source retrieval reproduced all feature fields exactly. Independent spectral and solver checks passed. A complete repeat matched all 90 analysis output files. The package preserves code, data, models, predictions, source locks, raw fixtures, figures, and reproduction instructions. Earlier paired results from the same patients informed this question. The test uses known event times within the same recordings. It does not establish continuous detection, motor causality, or performance in new sessions. Artifact masks were not applied.

stereoelectroencephalography · SEEG · high gamma · beta power · logistic regression · window classification · epilepsy · neural signals