An offline noise rule saved calibration pulses and raised mean SEEG forecast error in eight sessions

We tested an offline rule that stops SEEG calibration after three to five pulses at each stimulation site. We used OpenNeuro ds004696 version 1.0.1: eight patient sessions, 86 sites, and 886 bipolar readouts. The rule used 353 pulses instead of 430. It saved 17.9 percent of the pooled total and 20.5 percent on the equal-patient average. It stopped before five at 41 sites. Mean normalized error on later individual pulses rose from 0.63669 with five pulses to 0.65837. A random permutation of the chosen pulse counts held each patient pulse budget equal. Its exact expected mean error was 0.67312. The rule gained 0.01475 against this reference; the primary two-sided patient sign-flip p value was 0.0625. This gain remains uncertain. All fixed response-window checks saved pulses and raised mean error relative to five pulses. Test pulses followed calibration within each site, but site times interleaved across every session. Source-good pulse labels were retrospective. These results do not establish live interface performance. The package preserves the fixed local plan, exact matrices, all decisions, figures, source recipes, and code. Complete public source rebuilds matched three patients. A separate replay matched all numerical files and PNG figures. This is a new analysis of a previously used cohort, not new recordings or an independent replication.

SEEG · electrical stimulation · brain-computer interface · calibration · pulse allocation · forecast error · noise rule · offline analysis