After calibration and model fitting, two-bit power features gave 71.2% mean balanced accuracy in eight later-trial SEEG tests

We tested a fixed two-bit code for 24 extracted SEEG log-power features in all eight recordings from OpenNeuro ds004473 version 1.0.2. Each decoder used eight training-selected bipolar pairs. Encoder limits and model fits used only early trials. On 674 later trials, mean balanced accuracy was 71.2% with two-bit features and 70.7% with original float32 features. Each patient had equal weight. The paired change was +0.53 percentage points. Its 95% patient bootstrap interval was -0.87 to +1.85 points. The exact patient sign-flip test gave p=0.484375. These results do not prove equal accuracy. The feature message used six bytes instead of 96 bytes per window with shared configuration. The package preserves a reusable encoder, all features, models, predictions, packed messages, three figures, source versions, hashes, and retrieval code. Independent checks passed. A second analysis matched 27 numerical artifacts. Event times defined the windows. Quality screening used signal after the response. Earlier studies also used these test trials. This is an offline task-state test. No energy saving, continuous control, or brain stimulation effect was measured.

SEEG · neural feature quantization · brain-computer interfaces · balanced accuracy · feature compression · logistic regression · OpenNeuro