Clipped trial products reduced geometric mean squared forecast error for later 20–500 ms SEEG energy estimates in eight sessions

A fixed clipped trial-product estimator forecast later SEEG response energy with less error than squared mean voltage. We used public OpenNeuro ds004696 version 1.0.1 from eight epilepsy patient sessions. At each of 86 stimulation sites, the first five good pulses supplied forecasts and the final five supplied a signed trial-product target. The tested window was 20 < t < 500 ms. There were 886 bipolar readout edges and 430 unique later physical pulse events. Each edge loss was scaled by early mean single-pulse energy squared. Readouts, sites, and patients received nested equal weights. The geometric patient MSE ratio was 0.783, a 21.7% reduction; the patient bootstrap 95% interval was 0.653–0.925. Six patients had less error and two had more error. The exact two-sided patient sign-flip p value was 0.0390625 under independent, sign-symmetric patient effects. Clipped energy and energy of a scaled voltage forecast had almost equal error. Before stimulation, zero had less error than clipping. The result concerns a later measured energy estimate within a session. A stable mean and independent zero-mean trial noise are needed to interpret it as mean response energy. The study did not test neural origin, behavior, stimulation safety, or online BCI benefit. The package preserves trial products, all forecasts and losses, three figures, source hashes, code, and reproduction instructions.

stereoelectroencephalography · trial products · response energy · forecast error · electrical stimulation · brain-computer interface