Four bits per power feature gave 70.67% mean balanced accuracy in fixed SEEG tests on later trials from eight people

We tested feature precision for a reading brain-computer interface with public SEEG data. The source was OpenNeuro ds004473, version 1.0.2. Each of eight people had a separate logistic model with eight bipolar pairs and 24 power features. Models learned from earlier trials. They were tested on 1,348 pre-response and rest windows in later trials from the same recording. With the model fixed, four-bit inputs gave 70.67% mean balanced accuracy. Original inputs gave 70.69%. The mean change was -0.013 percentage points. Its 95% person bootstrap interval was -1.393 to +1.536 points. One person lost 3.488 points. Four-bit inputs changed 3.432% of decisions as an equal-person mean. A fixed score bound selected four bits, eight bits, or the original float32 inputs. All 1,348 adaptive decisions matched the reference model. Mean actual packet size was 167.86 bits, including a mode header and byte padding. This was 78.1% below 768 bits for the original features. This method assumes the same model at both ends and unchanged packet bits. The sender must compute the original score. The proof uses real arithmetic. These offline tests do not measure live control, power use, or stimulation. The package preserves code, exact inputs, model states, test decisions, packets, figures, raw-data checks, hashes, and retrieval instructions.

stereoelectroencephalography · brain-computer interface · feature quantization · balanced accuracy · adaptive precision · logistic regression · signal processing