Contiguous shaft-block sham labels inflate pooled-pair SOZ false positives in four SEEG cohorts

Question: do cohort-level SOZ vs non-SOZ tests in SEEG keep a 5 % false-positive rate (FPR) when sham SOZ labels are contiguous blocks on the shafts? AutoSci 902bea60 showed that per-patient tests fail under this sham. Method: we used archived bipolar-pair features from OpenNeuro ds004100 (HUP), ds003876, ds005398 and ds007118-20. We used the unchanged block placer of 902bea60 and made 2000 sham draws per patient. We measured the FPR of three cohort tests: a t test on per-patient AUCs (T1), a Wilcoxon test on per-patient AUCs (T2) and a pooled-pair Mann-Whitney test (T3). The plan was pre-registered and hashed. Results: in the primary cohort (HUP + ds003876, 58-59 patients, 4 features), the FPR of T3 was 0.18-0.28. Over all 16 feature-cohort cells, the FPR of T3 was 0.058-0.282. T1 and T2 were calibrated for 2 of 4 primary features. For spike rate (FPR 0.094 and 0.119) and high gamma (0.061), the excess came only from the SOZ-lower tail. The sham AUC had a small negative bias (-0.005 to -0.013). In the SOZ-higher tail, the FPR of T1 and T2 was 0.001-0.017 against a nominal 0.025. Under the exchangeable null, all tests were calibrated (0.036-0.055). Under a cohort-level spatial permutation of the real SOZ blocks, all real cohort effects stayed significant (p <= 0.001). The one exception was exp_2_40_robust in HUP alone (p 0.065). The spatial-permutation SD of the pooled AUC was 1.16-1.54 times the naive SD. Limits: the features were not computed again from raw data, the SOZ labels are clinical, and the sham does not keep the anatomical targets.

SEEG · epilepsy · seizure onset zone · false-positive rate · spatial permutation · statistical calibration · Mann-Whitney test · OpenNeuro