Gains from other stimulation sites increased normalized 20 to 500 ms SEEG forecast error relative to local gains in eight recordings

A gain is a number that scales a response template. We tested whether gains can transfer between stimulation sites in public SEEG data. We used OpenNeuro ds004696 version 1.0.1 and readouts selected from electrode metadata. Each forecast used the first five pulses at its target site. The transferred gain used five-pulse data from other sites at the same recording pair in the same person. All current settings at the target site were excluded. There were eight people, 314 stimulation sites, 317 conditions, and 1,183 condition–readout cases. For the 20 to 500 ms window, transfer increased normalized forecast error by a mean 1.86% relative to local gains. Error increased in all eight people. The descriptive 95% person bootstrap interval was 1.16% to 2.52%. The increase remained with source quality filters. The raw squared-error increase was 16.09%. Transfer still reduced normalized error by 7.30% relative to the unscaled mean. Its mean skill relative to zero was 7.10%, compared with 8.75% for local gains. A fresh raw rebuild used 192 hashed MEF files. It reproduced all saved inputs exactly. The package contains the plan, four figures, code, scores, source data products, and a raw retrieval recipe. This is an extension of a reused clinical cohort. It does not test a new person, fewer target pulses, or an online stimulation schedule.

stereoelectroencephalography · electrical stimulation · forecast error · gain transfer · calibration