Geometric medians of five traces increased 20–500 ms SEEG error compared with means in eight sessions without added faults

We tested geometric medians of five complete pulse traces for later SEEG voltage prediction. The public data came from OpenNeuro ds004696 version 1.0.1. We used 86 fixed readouts in eight clinical epilepsy sessions. Five early good pulses fitted each forecast; 612 later good pulses tested it. The window was strictly 20 to 500 ms after each pulse. On recorded inputs, the mean person percentage error increased by 3.24% compared with the ordinary mean. Its descriptive 95% interval was 1.26% to 5.56%. The fixed fault test changed only the third calibration trace. The added transient had four times the original calibration RMS, with both signs tested. Under that fault, mean person percentage error fell by 60.09%. Its interval was 53.95% to 67.60%. The scaled mean was better without added faults and worse under this large fault. A sample median had similar fault performance. A proof bounds median distance by 8r/3 when four inputs lie within radius r of a fixed center. All eight complete public source files passed fresh hash checks. All selected waveforms matched exactly. Independent objective and loss checks passed. A second run gave identical result files. The package preserves exact data, code, provenance, retrieval instructions, tables, and four figures. This offline test uses a reused cohort and retrospective quality labels. It does not measure clinical benefit or fault frequency.

stereoelectroencephalography · geometric median · pulse response · robust estimation · waveform prediction · fault tolerance · calibration