A fixed spectral model predicted random patient labels above chance across separate contacts and named shafts in 36 known SEEG patients

We tested random patient labels in public SEEG data from OpenNeuro ds004100, version 1.1.3. The test used 4,252 contacts and two 120-second records from each of 36 patients. A fixed random forest used four spectral measurements and equal patient training weights. Across 128 label draws, mean accuracy was 69.45% with separate contacts. It was 66.04% with separate named shafts and 50.05% with separate patients. The contact minus patient difference was 19.40 percentage points. Its 95% Monte Carlo mean interval was 18.82 to 19.98 points. These intervals describe random label uncertainty in the fixed cohort. They do not describe new patients. Separate records, common average inputs, and random contact labels supplied planned checks. All 72 raw prefix hashes matched. Rebuilt features agreed within numeric tolerance. All primary fits repeated exactly. The result shows that patient-linked structure can support high contact accuracy without a clinical target. The physical source of this structure and its effect on clinical models remain untested. The package preserves source features, labels, splits, predictions, code, figures, checks, and retrieval instructions.

SEEG · random labels · patient effects · data leakage · evaluation splits · random forest · spectral features