Linear maps fitted to pulse responses had more 20 to 100 ms SEEG forecast error than background maps in five selected sessions at ridge strength 0.1

We tested two matched linear maps on public stereoelectroencephalography (SEEG) data from OpenNeuro ds004457, snapshot 1.0.2. One map used pulse responses for training. The other used voltage before the pulse. Both maps used the same 603 earlier pulses, feature count, local four-pulse calibration and fixed ridge strength of 0.1. They forecast centered voltage at 20 to 100 ms on 268 separate pulses at 67 later sites in five selected patient sessions. The response map had more error in every session. The mean patient error increase was 57.79 percent. The median was 10.70 percent. One patient had a 257.25 percent increase. The background map also had more error than a four-pulse template in four sessions. Stronger penalties and recording pairs with no shared contact changed the result. These checks limit the claim to the tested model and split. The package includes a public benchmark, source hashes, code, pulse roles, fitted maps, per-pulse errors and three figures. All 110 raw source files passed hash checks. A new extraction matched every input array exactly. An independent solver checked 40 maps and 240 forecast sets. This offline test does not establish a causal brain-state effect or BCI performance.

stereoelectroencephalography · electrical stimulation · linear prediction · ridge regression · brain-computer interface