Across 117 SEEG patients in four cohorts, SOZ AUC spread exceeded simulated sampling noise (Q p = 0.040); exploratory true spread was 46%

We asked whether between-patient variation in interictal seizure-onset-zone (SOZ) area under the ROC curve (AUC) exceeded within-patient sampling noise. We analyzed archived per-pair feature tables from four public SEEG cohorts (117 patients), using the same rows and primary feature as the source AutoSci contribution and no new signal processing. For each eligible patient (at least five pairs, including at least one SOZ and one non-SOZ pair), we calculated AUC and Hanley–McNeil variance. The primary test summed cohort-specific Cochran Q statistics and compared the result with 2,000 binormal simulations under a no-heterogeneity null, preserving each patient’s pair counts. The observed Q was 590.6; the simulated null mean was 247.0, its 95th percentile was 539.5, and p = 0.040. Holm-adjusted cohort p values were 0.014 (HUP), 0.034 (ds003876), 0.11 (ds007118–20 NREM), and 0.11 (UCLA). The analytic Q calibration was invalid: under the simulated null, Q averaged about twice its 113 degrees of freedom, because Hanley variance was too small with few SOZ pairs or AUC near 1. DerSimonian–Laird I² (0.67–0.88) and reliability therefore overstated heterogeneity. Shaft-cluster bootstrap variance was lower than Hanley variance (median ratios 0.70 and 0.82 for HUP and ds003876), and did not indicate greater sampling noise from spatial correlation. An exploratory analysis added after the first run estimated that 46% of within-cohort AUC variance exceeded simulated sampling noise (pooled p = 0.0005); estimates by cohort were 67%, 45%, 40%, and 27% (UCLA p = 0.18). Under this calibration, even a perfect patient-level predictor could correlate with observed AUC only at r = 0.52–0.82 (about 0.68 pooled); a predictor explaining only part of true spread would have a smaller correlation. These are upper limits because spatial correlation may increase sampling noise beyond the simulation. Clinical SOZ labels are not ground truth; the null assumes binormal, equal-variance features and independent pairs. Only one feature was assessed per cohort.

intracranial EEG · SEEG · epilepsy · seizure onset zone · AUC · sampling noise · patient-level prediction · heterogeneity