High gamma to beta ratios reduced later trial loss against a constant model in eight SEEG patients

We tested whether high gamma to beta power ratios retain class information after common voltage gain is removed. We used public OpenNeuro ds004473, version 1.0.1. Each of 2,030 trials from eight SEEG patients supplied two half-second windows. One window preceded a key press. The other preceded fixation. Early trials trained each patient model. Later trials tested it. The fixed ratio model had lower loss than a constant probability 0.5 model in all eight patients. Equal-patient mean loss was 0.6093 versus 0.6931 nats per window. Mean gain was 0.0839 nats. Its descriptive 95% patient bootstrap interval was 0.0386 to 0.1358. The exploratory sign-flip p value was 0.0078125 under independent patients and symmetric null gains. Mean balanced accuracy was 68.38%. A full ratio and sum model gave loss 0.5761 and accuracy 70.26%. Removal of the sum therefore cost 0.0332 nats on average. With training constants fixed, ratio scores stayed fixed under positive gains within each window and electrode shaft. Each gain was constant within its window. Exact adverse gain tests measured the full model's sensitivity. These tests are mathematical envelopes, not observed error rates. All fixed checks had positive mean ratio gains. A weaker penalty gave gains in only five patients. Fresh pinned public byte ranges matched the source hashes in every patient. Independent spectral, solver, gain, and repeat checks passed. The package preserves code, data, models, predictions, raw examples, source locks, and reproduction instructions. Prior work on the same cohort informed the question. The results apply to known event windows within the source sessions. They do not establish continuous detection, motor causality, or transfer to new sessions. Artifact masks were not applied.

stereo-EEG · high gamma · beta power · power ratio · voltage gain · classification loss · within-session evaluation · event windows