Loss of packets with large voltage changes did not show a clear mean error rise over mask reassignment with linear repair and intact training in eight reused SEEG records at 20 percent loss

We tested whether loss that depends on SEEG voltage adds reader error beyond the layout of the missing samples. We reused eight records from public OpenNeuro ds004473, version 1.0.2. Each 500-sample window had eight bipolar pairs. The main rule removed five of 25 packets with the largest scaled squared voltage change. Twenty controls reassigned the same complete masks to other windows within each split. For linear repair and training on intact data, the mean Brier error difference was +0.002975. The descriptive 95 percent person interval was [-0.002711, +0.009407]. Four records had higher error and four had lower error. Training with selected loss reduced mean error from 0.211511 to 0.191915. Training with reassigned masks gave 0.189378. These training comparisons were secondary. The package preserves verified voltage buffers, mask donors, source code, models, scores, figures, source versions, hashes, and repeat instructions. This is an offline loss simulation on a reused clinical cohort. It does not test stimulation or a live BCI.

stereoelectroencephalography · packet loss · missing data · signal repair · Brier error · brain-computer interface