Window splits reduced exact squared risk from 0.362 to 0.091 for fixed five-neighbor closed-eye EEG predictions in 49 ds003490 people with 25 arbitrary positive labels

This study tested whether EEG validation can recover arbitrary person labels. The data were 14 closed-eye windows from each of 49 ds003490 people. A fixed five-neighbor model used posterior log-power features and 20 repeated five-fold splits. Labels were uniform among assignments with exactly 25 labels equal to one. Exact mean squared risk was 0.091454 with window splits and 0.362127 with person splits. In 10,000 label assignments, mean balanced accuracy was 88.49% and 48.55%. Eight planned checks retained the risk difference. The package contains a proof, a reusable exact-risk function, spectra, features, splits, neighbors, labels, source recipes, and reproduction commands. Independent numerical checks and an offline repeat passed. Fresh raw spectra from one control and one patient visit matched the archived spectra exactly. This result concerns shared information within recordings. It does not validate a disease marker or estimate risk in a new cohort.

EEG · validation leakage · person-level splits · fixed-count labels · squared risk · nearest neighbors · identity confounding · Parkinson disease