Signed power changes reduced mean response phase log loss with early setting selection in eight offline SEEG tests

We tested whether signed power changes help separate the periods before and after a recorded response. We used eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Readers used eight bipolar pairs and 500-sample windows. Early trials selected pairs, scales, and penalty settings. In 674 later trials, adding signed half-window power changes reduced equal-person mean log loss from 0.5471 to 0.5196 nats per window. Six of eight people improved. The paired mean reduction was 0.0275. Its descriptive 95% person bootstrap interval was 0.0019 to 0.0555. The exact sign-flip reference gave p = 0.1016. Mean balanced accuracy was 73.99%. The result depended on penalty strength: adding the same features at fixed C = 1 raised mean loss. Under symmetric Hann windows and mean removal, full waveform reversal preserves the unsigned features and negates signed changes. Reversing test changes raised mean loss in the fixed ordered reader. This check measures reader use of direction. It does not prove a neural cause. This exploratory test uses a reused clinical cohort and response-aligned windows. Cue signals can contribute. It does not test live control or stimulation. The package preserves fresh half powers, parent full powers, voltage examples, source versions and hashes, all models and predictions, code, three figures, and repeat instructions. Independent spectra, all 224 fits, all 12,132 predictions, and a complete numerical replay passed their checks.

stereoelectroencephalography · brain-computer interfaces · response phase · power features · logistic regression · offline evaluation