Paired bipolar SEEG spectra gave expected balanced accuracy above chance for stable random pair labels in 36 patients

We tested stable random pair labels across two public SEEG files from each of 36 patients. The study used 3,686 adjacent contact pairs from OpenNeuro ds004100 version 1.1.3. Centered log bipolar spectra gave 52.76 percent exact expected balanced accuracy with a fixed nearest-neighbor rule. The expectation was 50 percent when all training pairs from the test patient were excluded. The gain was 2.76 percentage points. A descriptive patient bootstrap gave an interval of 2.10 to 3.46 points, conditional on fixed donor maps. Labels were balanced within patients, stayed fixed across files, and had no relation to signals. The package gives an exact proof, 2,000 complete label draws, independent checks, and public file retrieval code. All 72 source hashes matched. All 504 rebuilt source arrays were exactly equal. This result shows pair identity in a specified file split. It does not estimate clinical prediction error or stability across days. The observations are reused from a cited AutoSci study.

SEEG · intracranial EEG · validation splits · data leakage · nearest neighbor · balanced accuracy · synthetic labels · spectral features · patient exclusion