The onset ranking loss gain interval included zero for added 20–100 Hz curvature in 15 SEEG test patients under fixed training and independent, exchangeable patient sampling

We tested if repeatable spectral curvature adds seizure onset ranking information in different patients. We used two first-minute SEEG files per patient from public OpenNeuro ds004100 version 1.1.3. The source cohort had 36 patients with epilepsy. The label rule retained 31 patients with both pair classes on at least one shaft. Fixed pairwise models used file 01 from 16 training patients. The primary test used file 02 from 15 other patients. The baseline used contact order, band power, and band slope. The added model also used curvature at 20–100 Hz, with 55–65 Hz removed. Mean loss fell from 0.253716 to 0.242885 nats per comparison. The gain was 0.010830. Its 95 percent patient bootstrap interval was −0.008537 to 0.027716. The interval assumes independent, exchangeable test patients and fixed training. It included zero. The primary evidence left the added mean ranking benefit uncertain. The median curvature map correlation between files was 0.908 across 36 patients. All five fixed sensitivity checks are reported. This is a secondary analysis of reused clinical labels and patients. The package preserves exact spectra, code, scores, source pins, raw examples, and reproduction instructions. Fresh raw reconstruction checked all 72 files. Independent numerical and raw audits passed. An offline repeat matched all numerical outputs.

stereoelectroencephalography · seizure onset · spectral curvature · pairwise ranking · patient bootstrap · epilepsy