Endpoint repair reduced mean 70 to 110 Hz power error in the first 100 ms after simulated 50 ms gaps in 21 SEEG patients with unchanged endpoints

A complete state repair reduced early SEEG band power error after known missing input intervals. We tested 35 public records from 21 patients in OpenNeuro ds006519 version 1.0.0. The main test used simulated 50 ms gaps, unchanged recorded endpoints, and a causal 70 to 110 Hz power feature. It used a past-only 100 ms mean of squared filtered voltage. Straight-line endpoint estimates were replayed through the band filter and the power mean. Over the first 100 ms after input returned, the equal-patient mean normalized RMS error fell from 0.7425 with frozen states to 0.3259 with complete repair. This was a 56.1 percent mean reduction. Twenty of 21 patients had less error. The paired mean reduction was 0.4166, with a descriptive 95 percent patient bootstrap interval from 0.2386 to 0.7052. Repair still underestimated early power in every patient. Later error persisted longer in the 13 to 30 Hz band. A finite-history proof and all 2,200 band-gap checks showed why both processing stages matter. The full rerun matched the numerical outputs. All selected inputs rebuilt exactly from fresh source files. The package preserves the repair tool, selected voltage, fixed plan, source recipes, score tables, and four figures. The evidence concerns signal processing after known gaps. It does not establish neural decoding benefit or safe stimulation timing.

SEEG · signal processing · missing data · causal filtering · band power · state repair