Power ratios kept 66.9% mean balanced accuracy under simulated constant ±6 dB bipolar channel gains in eight offline SEEG tests

We tested band power ratios in eight offline SEEG recordings from OpenNeuro ds004473 version 1.0.2. Both primary decoders used the same eight bipolar pairs and learned from earlier trials. They classified 674 later trials as pre-response or rest. Ratio features gave 66.9% mean balanced accuracy on unchanged signals. Absolute powers gave 70.7%. The paired accuracy cost was 3.8 percentage points. We then simulated constant broadband gains after bipolar contact subtraction. Each selected channel had one gain between -6 and +6 dB across all later windows. An exact search used later labels only to assess the minimum accuracy. Absolute models gave 49.8% mean accuracy at this minimum. Ratio models kept 66.9%. The primary paired difference was +17.1 points, with a 95% person-bootstrap interval from +10.1 to +25.3. The exact two-sided sign-flip p value was 0.0078125. Constant gains kept absolute-model score ranking unchanged. A known offset correction restored its predictions. The result concerns fixed thresholds under the stated gain model. It does not establish natural gain changes, continuous intention detection, or transfer across days. The package includes code, source records, trial features, raw audit ranges, fitted models, predictions, gain vectors, figures, and reproduction instructions.

stereoelectroencephalography · SEEG · brain-computer interface · band power ratios · bipolar signals · gain robustness · balanced accuracy