Fixed shared pulse weights preserved linear readout changes and gave no clear 20–500 ms SEEG forecast gain in eight sessions

We tested shared pulse weights for later SEEG response forecasts in eight reused patient sessions. The public source was OpenNeuro ds004696 version 1.0.1. The test had 317 conditions and 1183 condition/readout cases. Five early pulses fit each forecast. All later pulses tested it. At 20 to 500 ms, the mean person-level error change was +0.093 percent against the equal mean. The descriptive 95 percent person interval was -0.213 to +0.385 percent. This test showed no clear error reduction. Shared weights had 0.444 percent less mean error than separate readout weights. All eight people had less error. The exact loss decomposition attributed most of this difference to more equal weights. A proof showed consistency under any fixed linear readout change when the shared weights stayed fixed. Voltage examples showed defects when weights were recalculated after the change. The equal mean also preserves consistency. The package preserves inputs, code, all weights, errors, source pins, three figures, and reproduction instructions. A fresh source audit and raw-pair decode passed. The result does not establish a clinical BCI benefit.

SEEG · electrical stimulation · shared pulse weights · linear readout · forecast error · brain-computer interface · OpenNeuro