Finite SEEG contact counts account for about half of within-cohort variation in per-patient HFA AUCs

This study used the archived per-patient AUC table of 10 public SEEG cohorts (OpenNeuro ds004100, ds003876, ds003498, ds007118-20, ds005398). It divided the within-cohort variance of each patient AUC into finite-sample noise (Hanley-McNeil variance from the numbers of SOZ and non-SOZ contacts) and true between-patient spread. For 80-150 Hz power (HFA), noise is 0.56 [95% CI 0.49, 0.74] of the variance. The true between-patient SD is 0.15, and the median reliability of one patient AUC is 0.52. Reliability 0.8 needs approx. 24 SOZ contacts; the median is 5. Line length and 1-40 Hz power are similar (0.52, 0.62). The aperiodic exponent and spike rate have less noise (approx. 0.35). Patients with fewer SOZ contacts have AUCs further from 0.5 (HFA rho -0.29, p = 0.002). The patient-level inverse-variance analysis S1 in the source meta-analysis is controlled by 5 patients with extreme AUCs (55% of the weight). Patient-level moderator tests on raw AUCs are attenuated; use shrinkage or weights by contact number. The contacts are correlated, so these noise fractions are probably lower bounds.

stereo-EEG · seizure-onset zone · area under the curve · finite-sample variance · measurement reliability · high-frequency activity · epilepsy · patient-level analysis