Removing contact means from 32 early baseline windows raised mean later balanced accuracy by 6.34 percentage points over a raw median in eight SEEG recordings

We tested removal of early contact means before a median reference in public OpenNeuro ds004473 version 1.0.1. Eight epilepsy recordings supplied 1864 retained trials, with 1300 earlier training trials and 564 later tests. Each contact mean used the first 32 baseline windows. Each reference had a separate ridge reader on the same eight training-selected contacts. Clean mean balanced accuracy was 64.13% after correction and 57.79% with the raw median. The fixed paired difference was +6.34 percentage points; its descriptive 95% person bootstrap interval was +1.63 to +10.49. The sign-flip value was 0.0469. The corrected median gave 64.10% accuracy under white noise at 100 times the training scale on one unused contact; the raw median gave 57.91%. A proof shows invariance to contact constants present in both calibration and later use. The package preserves exact features, raw byte examples, fitted readers, predictions, code, source hashes, figures, and repeat instructions. This is an offline test with reused data and known task times. It does not test continuous intent reading or brain stimulation.

stereoelectroencephalography · brain-computer interface · median reference · contact mean removal · signal processing · balanced accuracy · noise robustness