Stored float32 packet means kept causal 70–110 Hz power errors below 5 percent of baseline for 100 ms packets in 21 SEEG patients

We tested whether packet mean removal preserves a causal SEEG power value. We used 35 public records from 21 patients in OpenNeuro ds006519 version 1.0.0. Each record supplied four fixed channels. For 100 ms packets, the patient mean rate of post-window power errors above 5 percent of earlier baseline was 79.2 percent. Its descriptive 95 percent interval was 76.2 to 82.3 percent. The later pre-window rate was 75.4 percent. A controlled input with only 1 to 30 Hz Fourier components also produced large 70 to 110 Hz power after packet mean removal. Storing float32 residuals and one float32 mean per channel per packet kept all primary readings below the error limit. The largest post-window error was 0.000685 percent of baseline. Mean storage added 0.244 percent of the float32 residual payload at 4096 Hz and 1.961 percent at 512 Hz. These costs exclude transport headers and device energy. The package supplies selected voltage, source metadata, a retrieval recipe, code, all scores, and three figures. Fresh source downloads rebuilt all 35 selected arrays exactly. Independent calculations passed, and a full repeat gave identical numerical results. This is a processing accuracy benchmark. It does not test a decoder or a live stimulation system. The patients are shared with earlier sci studies.

SEEG · brain-computer interface · packet mean removal · causal filtering · high-gamma power · signal processing · float32