A fixed EEG model had mean random-label balanced accuracy near 33% across rest and light stimulation in 72 shared people

We tested whether random person labels transfer between rest EEG and light-stimulation EEG. Both public datasets contain the same 88 people. A fixed window rule and a positive-power rule retained 72 paired people. The model used 95 log relative band powers and three nearest neighbors. Scaling used source training data only. Each comparison kept 71 training people and matched the supplied diagnosis counts. For 999 random label assignments, shared-person mean balanced accuracy was 33.8% from rest to stimulation and 33.4% in reverse. The chance reference was 33.3%. Separate-person means were 32.3% and 31.8%. Identity prediction found 2 of 72 people and 1 of 72 people. Controls added after the primary result gave within-condition shared-person means of 68.0% for rest and 52.7% for stimulation. Different epoch counts limit causal comparisons between tests. These results show little random-label transfer for this model and eligible cohort. They do not prove that all person information is absent. The package preserves features, exact predictions, source hashes, retrieval recipes, code, local plans, amendments, and checks. Fresh source replay reproduced checked features. Two numerical runs gave identical files.

electroencephalography · EEG · person-level data leakage · random-label prediction · cross-condition transfer · balanced accuracy · nearest neighbors