A three-neighbor EEG model gives 33.5% mean balanced accuracy for labels shuffled within folds across two tasks in 72 shared people

This study tested a fixed model on paired resting and photic EEG from OpenNeuro ds004504 and ds006036. The tested cohort had 72 people from one clinical session. Each task supplied four epochs per person. The model used 95 relative band powers and three nearest neighbors. Two training rules had equal sizes and original group counts. One rule included all test people. The other excluded them. Across 999 participant-label assignments shuffled within fixed folds, mean person balanced accuracy was 33.52% with shared people and 33.42% with separate people. The mean paired gain was +0.10 percentage points. Its central 95% assignment range was -9.65 to +9.35 points. Ten planned split repeats gave small positive mean gains from +0.10 to +1.22 points. A control chosen after the primary result used disjoint early and late resting epochs from the same people. It gave 55.87% with shared people and 33.06% with separate people. This evidence shows that a low shuffle score across tasks cannot confirm separate cohorts. It does not show that other models lack identity information or that task caused the score change. Fifteen sources had too few valid epochs. One further source had a selected epoch with zeros in all channels. A power exclusion was added before scoring. The package preserves this change, the inspected source file, features, all predictions, code, source versions and hashes, figures, and reproduction steps. Fresh main and control analyses gave identical output bytes. The result applies to the fixed model and eligible people.

electroencephalography · EEG · participant-label control · balanced accuracy · nearest neighbors · cross-task prediction