A fixed covariance fit had 6.6 percent more mean squared log error than uniform fits at 20 to 500 ms in six reused SEEG sessions

We tested whether uncertainty from five early pulses improved later SEEG response forecasts. We used public OpenNeuro ds004696 v1.0.1 products in six reused sessions, with 167 reciprocal contact pairs. All numerical fits used earlier pulses and omitted the tested pair. The shared-pulse rule used jackknife covariance plus a fixed median variance term. At 20 to 500 ms, equal-person mean squared log RMS error was 1.442, against 1.353 for uniform fits. The covariance fit had 6.6 percent more error. Five of six people had more error. Response gain minus before-pulse control gain was -0.111. Its descriptive 95 percent person bootstrap interval was -0.243 to 0.000210. The two-sided person sign test gave p = 0.21875. The separate-variance fit had error 1.391. This fixed uncertainty rule did not reduce mean response error in these sessions. Small numbers, reused targets, and retrospective source selection limit the result. The result does not establish that all covariance fits fail. The package includes code, pulse products, covariance arrays, donor weights, three figures, and source provenance. One complete processed source file passed hash and voltage array checks. A repeated analysis gave identical output bytes.

stereoelectroencephalography · brain stimulation · response forecasting · graph models · jackknife covariance · calibration