A fixed EEG model predicts random person labels across one month when training includes each test person in 60 ds004148 v1.0.1 adults

A split between EEG sessions did not remove person information in 60 healthy adults from ds004148 version 1.0.1. Training used session 1. Testing used session 3, one month later under the source protocol. A fixed nearest-neighbor model used 28 posterior log-power features. We gave each person a random binary label that stayed fixed across sessions. Across 9,999 balanced label assignments, mean accuracy was 76.34% when training included the test person and 49.20% when it excluded that person. Both models used 59 training records per prediction. The exact expected difference was 27.12 percentage points, explained by 32 own-person neighbors in 60 tests. Independent test-session labels gave mean scores near 50%. This test does not measure disease prediction or identify a brain mechanism. The package preserves all 360 input spectra, source metadata, code, random labels, results, and reproduction instructions. Separate numerical checks passed. All nine analysis outputs had identical hashes after a repeat. Source replay rebuilt all 120 primary records with exact array equality.

EEG · subject leakage · cross-session prediction · random-label control · nearest neighbor · validation · spectral power