Bootstrap copy splits gave 77.30% mean random-label accuracy in five-fold one-neighbor EEG tests of 88 ds004504 v1.0.9 people

We tested EEG cross-validation after sampling people with replacement. We used one fixed feature vector for each of 88 people in OpenNeuro ds004504 v1.0.9. Each of 9999 draws assigned balanced random labels to the original people. Splitting sampled copies between five folds gave 77.30% mean one-neighbor accuracy. The exact expected mean is 77.38% when features, folds, scaling, and neighbor choice do not use labels. Removing all training copies of the original test people gave 49.27%. Keeping each original person in one fold gave 49.36%. The exact exclusion reference is 49.43%. The package gives the proof, a three-neighbor check, all sampled IDs and predictions, and an original-ID guard. Three complete runs gave identical numerical files. A new source-file check rebuilt one person's 20 EEG feature rows with exact agreement. The package preserves code, source hashes, results, and reproduction steps. This result tests a known copy problem under fixed rules. It does not establish an EEG relation to disease or clinical performance.

bootstrap cross-validation · duplicate observations · data leakage · EEG · nearest neighbors · random labels · identity-aware validation