A numpy-only pipeline computes per-patient SOZ-versus-non-SOZ AUCs of 16 interictal SEEG features in OpenNeuro ds004100; in a 5-patient pilot, no feature passed Holm correction

This archived pilot asks whether 16 simple features from 5-minute interictal stereo-EEG (SEEG) clips distinguish clinically marked seizure-onset-zone (SOZ) contacts from other contacts in the HUP iEEG Epilepsy Dataset, and whether results change by reference, shaft location, surgical outcome, or clip. A numpy-only pipeline streamed EDF files from OpenNeuro ds004100 snapshot 1.1.3 and recorded SHA-256 hashes. It analysed bipolar and common-average references, spectral features, aperiodic exponents, line length, and a transient rate. Only 5 of 38 eligible participants were processed because some S3 downloads failed with SSL errors and others exceeded the 45-minute run limit. For run 01 with bipolar reference, log power in all six bands and line length had AUC above 0.5 in all five patients; mean AUCs ranged from 0.639 to 0.724. Relative high-gamma power had mean AUC 0.334. No feature passed Holm correction; with five patients, the smallest possible sign-flip p-value was 0.0625. Run 02 had similar mean AUCs for log power (0.585–0.647), while run-to-run patient AUC correlations varied by feature. Only one patient met the same-shaft comparison criterion, and outcome groups contained two and three patients. These results have no inferential value at this sample size and do not establish whether any feature localizes the SOZ. The archived report identifies completing the 38-patient cohort with lower download parallelism as the next step.

stereo-EEG · seizure-onset zone · epilepsy · interictal · AUC · OpenNeuro · pilot study