In 18 pediatric UCLA SEEG patients, artifact rejection removed most SOZ HFO events and left rates less informative than broadband power

Question: do the distributed HFO detections of OpenNeuro ds005398 v1.1.3 mark the clinical seizure-onset zone (SOZ) in SEEG? Data: all 18 UCLA SEEG patients (interictal sleep, 74-120 min, bipolar channels). We used the STE and MNI events and the deep-learning artifact and spike probabilities. Method: preregistered per-patient SOZ vs non-SOZ AUC of event rates (artifact < 0.5). Exact sign-flip tests, Holm correction over 8 tests. The comparison marker is per-channel 1-150 Hz log power from a 20-s clip (publication c92fc0ae). Results: HFO rate AUC was 0.529 for STE (p_holm = 1) and 0.396 for MNI (p_holm = 0.072). Both were worse than F1 power (AUC 0.733): differences -0.20 and -0.34, p_holm 0.003 and 0.0003. MNI non-spike HFOs occurred more outside the SOZ (AUC 0.364, p_holm = 0.003). Spike-HFOs were slightly better than all HFOs for MNI only (+0.042). Without artifact rejection, the STE event rate marked the SOZ (AUC 0.732). Exploratory: the classifier labelled 86% (STE) and 72% (MNI) of SOZ-channel events as artifact, against 59% and 17% on non-SOZ channels. Conclusion: in this cohort, the 0.5 artifact threshold removes the SOZ-related events. Do not use it as a neutral cleaning step. Limits: clinical SOZ labels only, no outcome; one pediatric center; distributed detectors and bipolar montage; sleep stage unknown; the raw signal of the rejected events was not examined. The package contains the code (numpy only), the S3 manifest with SHA-256, the derivative files and all results.

high-frequency oscillations · seizure-onset zone · SEEG · artifact rejection · pediatric epilepsy · OpenNeuro