Shaft clustering inflated pair-level SOZ test positives in 87 patients across four SEEG cohorts

In 87 patients from four public SEEG cohorts, naive pair-level Mann–Whitney tests produced false positives for a mean of 19.6% of random shaft-clustered labels (95% bootstrap CI 17.2–22.1%; 82/87 patients exceeded 5%). A block null that moved whole SOZ runs among random shaft positions produced a 3.8% rate (95% CI 3.3–4.3%), slightly below the nominal 5%. For clinical SOZ labels, 46/87 patients showed a significant higher feature value under the naive test and 24/87 under the block null; 22 were significant only under the naive test, and none only under the block null (McNemar p = 2.4×10⁻⁷; Holm-adjusted p = 4.8×10⁻⁷). Median between-shaft feature variance share was 0.31, and its exploratory Spearman correlation with the naive false-positive rate was 0.42. The block null preserved SOZ run lengths but did not model electrode placement or relations between shafts. Analyses used short archived interictal clips and one feature per cohort; clinical SOZ labels are not ground truth.

stereo-EEG · SEEG · epilepsy · seizure onset zone · SOZ · false positive rate · Mann–Whitney test · block permutation · electrode shafts · interictal biomarkers