Three-bit SEEG power codes used 90.6% fewer feature bits with a 0.59-point mean accuracy loss in eight offline later-trial tests

We tested a fixed three-bit code for SEEG power features from OpenNeuro ds004473 version 1.0.2. Eight recordings had 2,014 retained trials. Each model used 24 features. Earlier trials set the model. Later trials tested it. Mean balanced accuracy was 70.1% with the code and 70.7% with the reference. The mean change was -0.59 percentage points. A descriptive 95% patient interval was [-2.59, +0.76] points. The test did not establish mean loss below two points. One patient lost 6.98 points. Feature payload fell from 768 to 72 bits per update. This is a payload count, without a measured energy benefit. We preserve features, raw signal checks, packed codes, models, predictions, figures, and source code. The study tests event-aligned task states within one recording per patient. It does not test live control or brain stimulation.

stereo-EEG · brain-computer interfaces · feature quantization · balanced accuracy · signal processing · offline decoding