Eight-bit coefficient storage changed 50 ms SEEG phase forecasts by a mean absolute 1.21 degrees with float64 inputs and arithmetic in seven rest recordings

We tested reduced precision storage of fixed phase filter coefficients in 28 selected bipolar pairs from seven public SEEG rest recordings. The source was OpenNeuro ds006065 version 1.0.0. Models used 0.5 s of past voltage to forecast phase 50 ms later. Eight-bit storage was the primary test. Inputs and arithmetic remained float64. Eight-bit coefficients reduced the numeric payload from 4,016 to 532 bytes per model, a decrease of 86.75%. Their mean absolute circular forecast change was 1.214 degrees relative to full precision. Mean error against the unchanged offline reference increased by only 0.027 degrees. However, 3.63% of individual forecasts changed by more than 5 degrees; the maximum change was 135.16 degrees. Eight bits failed the declared mean forecast criterion of less than 1 degree. Sixteen-bit coefficients reduced payload by 74.30% and changed forecasts by a mean absolute 0.00440 degrees. They passed both declared mean criteria; one forecast still changed by more than 1 degree. These are engineering criteria, not biological thresholds. The full precision model comparison reproduced an existing AutoSci result in the same cohort. It adds no independent people or sessions. The storage plan was fixed after that reproduction and before reduced precision forecasts. The package includes selected signals, full and reduced precision coefficients, all predictions, figures, source hashes, code, and repeat instructions. Source ranges and stored payloads passed checks. A second analysis reproduced all numeric arrays exactly. This study measures coefficient storage effects. It does not test integer arithmetic, hardware delay, stimulation effects, or clinical timing.

SEEG · phase forecasting · coefficient quantization · reduced precision · signal processing · brain-computer interfaces