Band power ratios retain 67.0% balanced accuracy under simulated constant voltage gains per shaft in eight SEEG recordings

We tested whether band power ratios preserve small SEEG decoders during simulated recording gain changes. We used OpenNeuro ds004473, version 1.0.2, with one recording from each of eight patients. The models used eight bipolar pairs and three power bands. They learned from early trials and classified rest and pre-response windows in 674 later trials. Each shaft had one voltage gain, shared by its contacts and constant over time and frequency. At gains from 0.5 to 2, pair ratios retained 67.0% mean balanced accuracy. Absolute power gave 70.7% without a gain change and 49.8% at its exact minimum over allowed gains. The planned patient contrast was 17.3 percentage points, with a 95% patient bootstrap interval of 10.5 to 25.3 points. The exact two-sided sign-flip p value was 0.0078125. Pair ratios had a 3.6-point lower mean accuracy without a gain change. A planned shaft ratio model retained 69.0%. All models retained their score ranking under constant gains. The result concerns a fixed decision threshold and the stated fault set. It does not establish live control or performance during real hardware faults. The package preserves features, raw audit ranges, source versions, code, fitted models, predictions, figures, and retrieval instructions.

SEEG · band power · power ratios · gain robustness · balanced accuracy · brain-computer interface · signal processing · logistic regression