SEEG spectra with level and slope removed at 20–100 Hz gave 65.8 percent expected accuracy for shared balanced random pair labels across two files per patient in 36 patients

We tested whether residual spectral shape can transfer arbitrary pair labels between two public interictal SEEG files. We used OpenNeuro ds004100 version 1.1.3 and 36 patients. Each pair had the same random binary label in both files. Labels were balanced within each shaft. We removed power level and linear slope from log spectra at 20–100 Hz, with 55–65 Hz excluded. Each query selected the nearest source spectrum on its shaft. Across 3,675 pairs, the exact expected label accuracy was 65.83 percent. Uniform source choice gave 51.14 percent. The mean gain was 14.69 percentage points, with a 95 percent patient bootstrap interval of 13.09 to 16.34 points. A stronger constant label baseline gave 56.10 percent. Its mean gain was 9.72 points, with an interval of 8.13 to 11.34 points. This second comparison was added after the planned results. The result shows a possible route for transfer of fixed pair labels across the supplied files. It does not test clinical labels, new patients, or independent recording sessions. The package preserves all input spectra, source identities, analysis code, patient results, the exact label formula, figures, and reproduction instructions. Eight fresh raw file checks covered all four sample rates. All planned conditions are reported.

stereoelectroencephalography · SEEG · spectral shape · random labels · label transfer · data leakage · signal processing · epilepsy