Released STE counts kept a higher mean SOZ AUC after at most two shaft omissions per patient that kept both classes in 18 first-hour SEEG records

We tested whole-shaft omission in 18 public SEEG sleep records from OpenNeuro ds005398 version 1.1.3. We used released STE and MNI event counts in the first hour. Each patient had equal weight. Retained channel counts and source SOZ labels stayed fixed. The same omission acted on both detectors. Both label classes had to remain. The mean STE minus MNI AUC was 0.201090 with no omission. Its exact range was 0.135641 to 0.257770 with at most one omitted shaft per patient. With at most two, the range was 0.081443 to 0.297024. Thus, the mean order did not reverse in the tested set. Six single-shaft omissions removed all recorded SOZ channels and had no AUC. Individual patients could be sensitive despite the cohort result. Half-hour checks and removal of UCLA42 kept positive bounds. An exact group-sum tool and budget algorithm give all endpoints and omission witnesses. Independent direct-pair checks passed. A fresh offline build reproduced nine numerical files byte for byte. Fresh downloads matched all 94 source files or header ranges. The package preserves source CSVs, metadata, EDF headers, hashes, code, proof, results, figures, and repeat instructions. This is a finite-cohort sensitivity test. It does not assess unsampled brain tissue, validate raw HFO signals, or establish clinical benefit.

SEEG · high-frequency oscillations · shaft omission · area under the ROC curve · sensitivity analysis