A geometric median of five early pulses gave an uncertain 0.4 percent decrease in later SEEG prediction error at 20 to 500 ms in eight patients

We tested whether a geometric median of five early SEEG maps could predict five later pulses better than their mean. Public data from OpenNeuro ds004696 version 1.0.1 gave eight epilepsy patients, 86 sites, and 886 bipolar readouts. At 20 to 500 ms, the equal-patient mean relative error decrease was 0.388 percent. The descriptive 95 percent patient bootstrap interval was -0.826 to 1.296 percent. The evidence does not establish a benefit on the recorded pulses. A fixed digital test multiplied the fifth early map by four and kept the original clean readout scales. Under that specified fault, the median gave a 35.197 percent mean error decrease. All 430 fits agreed with a separate Newton solver. A full analysis repeat matched all seven main output files. Archived inputs and three fresh public metadata files passed hash checks. A fresh source-file rebuild failed during download. The package records that failure and retains earlier source audits. This work provides an exact Gram calibration benchmark. It does not test BCI command decoding or clinical stimulation. The package contains code, exact data, provenance, all scores, and three figures.

stereoelectroencephalography · SEEG · geometric median · brain-computer interface · electrical stimulation · pulse-map prediction