4-bit SEEG power features met a mean loss limit of 2 percentage points in eight later trial tests with fixed training bounds and refitted models under paired t assumptions

We tested feature precision in public SEEG data from eight patients in OpenNeuro ds004473 version 1.0.2. Earlier trials set eight bipolar pairs, 24 power features, and fixed bounds. We refitted each model to quantized training features. In 674 later test trials, mean balanced accuracy was 71.30% with 4-bit features and 70.69% with float features. The paired mean change was +0.61 percentage points. Its two-sided 95% t interval was -0.14 to +1.37 percentage points. Under independent normal patient effects, the fixed mean loss rule of 2 percentage points passed. This result does not establish an accuracy gain. Each 24-feature window needs 12 packed bytes instead of 96 float32 bytes, plus 192 fixed bound bytes per patient. The package preserves code, source features, raw audit ranges, models, predictions, packed payloads, and three figures. The test concerns offline event-aligned states in the same recording. It does not measure continuous intent, transfer between days, device energy, or stimulation.

SEEG · stereoelectroencephalography · intracranial EEG · feature quantization · brain-computer interface · balanced accuracy · bit packing · signal processing