Four-bit SEEG features clipped at three training standard deviations gave 70.9% mean balanced accuracy with fixed decoders on later trials from eight patients

We tested whether low input precision changed a fixed SEEG decoder on later trials. We used public OpenNeuro ds004473 version 1.0.2 data from one recording in each of eight patients with epilepsy. Each decoder used 24 band-power features from eight bipolar pairs. Pair selection, scaling, and fitting used early trials. The primary encoder used four bits per feature and a range of three training standard deviations on each side of zero. The test had 1,348 decisions. Mean patient balanced accuracy was 70.86% with four bits and 70.69% with floating inputs. The paired change was +0.18 percentage points. Its descriptive 95% patient bootstrap interval was -1.04 to +1.16 points. Four bits changed 43 labels. The largest patient loss was 3.49 points. These results do not establish equal performance. Packed feature payloads used 12 bytes per window, compared with 96 bytes for float32 inputs. This eightfold reduction excludes other transmission data and feature computation. A code interval check certified preservation of the original decoder label for 53.47% of decisions with equal patient weight. A preserved label can still be wrong. The package contains source provenance, inherited features, pinned raw sample recipes, code, packed streams, all decisions, three figures, and an exact numerical rerun. The scope is offline task-state decoding within each recording. We did not test continuous control, hardware energy, or brain stimulation.

stereoelectroencephalography · brain-computer interfaces · feature quantization · low-precision decoding · balanced accuracy · label certification