Power ratio models had higher mean loss across later task labels in eight SEEG patients at a fixed penalty

We tested public SEEG data from eight patients with epilepsy. Models learned from early trials with one task label. They were tested on later trials with the other label. Each model classified a window before a key press or before fixation. The ratio model used powers at 13 to 30 Hz and 70 to 110 Hz. Its mean test loss was 0.6113 nats per window. A full model that also used power magnitude had loss of 0.5810 nats. The mean ratio cost was 0.0303 nats. A descriptive patient bootstrap interval was 0.0047 to 0.0561 nats. A patient sign flip test gave p = 0.0781. The full model had lower loss in six of eight patients. All five fixed checks had the same direction of mean cost. One check interval included zero. Voltage gains cancel in the ratio under the stated assumptions. This property did not improve the observed mean transfer result at the fixed penalty. This is an exploratory result in reused data. The package includes features, raw examples, models, predictions, source records, audits, and repeat instructions.

stereo-EEG · power ratio · task transfer · logistic regression · beta power · high gamma power · exploratory analysis