Across four SEEG cohorts, SOZ pairs had a narrower spread, explaining 83% of the top-pair shortfall under a binormal model

In archived interictal high-frequency activity (HFA) data from four SEEG cohorts, the top-ranked bipolar pair was in the clinical seizure-onset zone (SOZ) in 36 of 117 patients, compared with 55.3 expected under an equal-variance binormal model. We tested whether the ROC shape explains this shortfall using the same archived per-pair tables, row rules, and primary features as the prior analysis, with no new signal processing. Across 114 patients with at least two SOZ pairs (885 SOZ pairs), the pooled within-patient SD of SOZ-pair placement z-scores was 0.730, or 0.81 of the equal-variance null mean (0.905; one-sided p = 0.0005, the simulation floor). All four cohorts showed a narrower spread after Holm correction (R = 0.74–0.89; p = 0.002–0.032). A fitted common SOZ SD of 0.77 (95% CI 0.70–0.82; binormal slope about 1.3) predicted 39.4 top-ranked SOZ pairs and explained 83% of the shortfall (95% CI 63–122%); observed and fitted counts were not significantly different (p = 0.46). The narrower spread remained when one SOZ pair per shaft was selected (R = 0.86, p = 0.0005; 59 patients). The results support an asymmetric ROC shape as an explanation for why AUC overstates top-pair accuracy in these data. The analysis uses clinical SOZ labels and 90–120 s interictal clips across cohorts with different features. The binormal model does not distinguish a narrow SOZ distribution from a wide non-SOZ distribution, and between-shaft SOZ-pair correlation within a patient can also reduce the measured spread.

SEEG · epilepsy · seizure-onset zone · interictal high-frequency activity · ROC · binormal model · AUC · spread · electrodes · statistical methodology