Two 30 minute windows retained selection bias among four HFO scores under shared random channel labels in 18 SEEG records

We tested whether separate time windows remove bias from high frequency oscillation (HFO) score selection. We used public stereoelectroencephalography (SEEG) detector marks from OpenNeuro ds005398, version 1.1.3. There were 18 patients, 1,285 channels, and 181 source seizure onset zone labels. For each patient, we made 20,000 random label sets with the same positive count. We chose one of four count scores on the first 30 minutes. We tested that score on the next 30 minutes with the same labels. At a 5% test threshold, mean false rejection was 9.22%. It was 4.82% with independent later labels. It was 4.75% when each random reference assignment repeated score selection. The paired excess was 4.40 percentage points, with a 95% patient bootstrap interval of 3.85 to 4.97 points. The planned checks retained the excess. This result applies to fixed score maps and the specified artificial label model. It does not measure clinical error or prove patient benefit. The package has code, source files, hashes, per-draw results, and repeat instructions.

selection bias · permutation test · AUC · high frequency oscillations · stereoelectroencephalography · randomization inference · time split validation