A fixed posterior EEG rule predicts balanced random person labels across one-month visits in 60 shared ds004148 v1.0.1 adults

We tested random person labels across EEG visits one month apart. The source was OpenNeuro ds004148 version 1.0.1, with 60 adults. A fixed nearest-neighbor rule used posterior 1 to 40 Hz log power spectra with eyes closed. Each gallery contained 59 people. The shared gallery matched 23 of 60 queries to their own earlier record. For balanced random binary person labels, exact expected accuracy was 68.64% with the test person present. It was 49.15% with that person absent. The expected gain was 19.49 percentage points. All 9,999 sampled assignments gave a positive paired gain. The package preserves a proof, all 360 spectral records, nine planned descriptive checks, and all random draws. We verified all 120 complete source EEG records and checked FFT scaling with separate code. A full analysis rerun gave identical result bytes. The result applies to this fixed representation and cohort. It shows that separate visits can retain apparent prediction from shared people. It does not estimate clinical accuracy or performance in a new population.

electroencephalography · EEG · person identity · random labels · nearest neighbor · test-retest reliability · data leakage · balanced accuracy