32-bit biases raised mean site accuracy over 4-bit biases in five offline SEEG tests with 4-bit voltage weights at 20–100 ms

We tested whether stored site readers need more precise biases than voltage weights. We used OpenNeuro ds004457 version 1.0.2. Five patient sessions supplied 183 sites and 732 test pulses. At 20 to 100 ms, float32 biases raised mean site accuracy from 32.89% to 38.56% over four bit biases. Both readers used the same four bit voltage weights. Four patients improved. The added biases cost 618 numeric bytes across five readers. The 5,908 byte numeric payload was 82.86% smaller than the full float32 reference. The exact two sided patient sign flip p value was 0.125. Calibration corrections gave mixed gains at four bits. Recording time alone gave 74.80% mean accuracy in the same split. The results therefore describe coefficient storage in a fixed offline task. They do not establish a population benefit or a neural readout for live use. The package contains code, bins, raw retrieval hashes, fits, scores, predictions, packed readers, figures, and replay checks. All 110 raw source files passed hash checks. The rebuilt bin files matched the archived inputs.

stereoelectroencephalography · SEEG · coefficient quantization · bias precision · site classification · electrical stimulation · brain-computer interface