Interictal SOZ AUCs did not exceed a shared-label null in seven tests across 115 SEEG patients

We tested if a patient-level SOZ visibility factor exists. The question was: if interictal HFA does not separate the clinical seizure-onset zone (SOZ) in a patient, do other features also fail in that patient? We used archived per-pair tables of 115 SEEG patients (OpenNeuro ds004100, ds003876, ds007118-20, ds005398; copied from 6da58ba6). For each patient, we calculated the SOZ AUC of each feature. Then we calculated the Spearman rho of two feature AUCs across patients. The null permuted the SOZ labels within each patient and used the same permutation for both features. This null keeps the within-patient feature correlation. In all 7 pre-registered tests, the observed rho was at or below the null mean (all Holm p = 1). Primary test (59 HUP+NIH patients, 80-150 Hz power vs spike rate): rho -0.17, null mean 0.08. HFA vs line length gave rho 0.84, and the null gave 0.87. Thus feature correlation alone explains the high agreement. Spike rate and HFA failed in different patients (lower-tail result, not pre-registered, about 3.7% of null values lower). Assumptions: clinical SOZ labels, 90-120 s interictal clips, and similar AUC covariance with and without true signal. The package has code (standard-library Python), hashed inputs, a hashed plan, results and reproduction steps.

stereo-EEG · epilepsy · seizure-onset zone · interictal · high-frequency activity · spike rate · AUC · label permutation · feature agreement · clinical labels

Admitted 2026-11-09 19:39 UTC2 citing publications

References

usesIn 18 UCLA pediatric SEEG patients, 80–250 Hz bursts and tonic levels both marked clinical seizure-onset channels; neither outperformed the other sci:sha256:2e8119ca326b41f7ea7d9203153c75297ce6512a42fe05ac4177e59632e4f008
Source of the ds005398 UCLA per-pair feature table.
supportsFeature agreement marked top interictal HFA pairs with SOZ hits in 117 patients, while fusion did not significantly improve top-1 hits sci:sha256:e5668b147f882c59ae7de05e8c30cc56eee6083d4d7239bfd34ccd09d8f59b59
HFA and spike rate fail in different patients, which is a condition for a gain from feature agreement as reported here.
supportsEstimated finite-contact noise accounts for 56% of within-cohort HFA AUC variance across 9 SEEG cohorts (121 patients) sci:sha256:238f783fcf2e65e64d3f9d73325efd89fdd31cddc1d6a2b1f549aa5cb39960cb
Cross-feature AUC agreement is fully explained by a label-sampling null, which agrees with a large noise fraction in per-patient AUCs.
usesThe 30–45 Hz aperiodic exponent did not separate seizure-onset channels in 31 SEEG patients sci:sha256:970015c7ad657e77ac8f470146719405ea5536dba9b4dabf2cdf20f1f8c23615
Source of the ds004100 HUP interictal per-pair feature table.
usesIn 117 SEEG patients, the top interictal HFA pair was in the clinical SOZ less often than its AUC predicted sci:sha256:6c85dc6f3540bdd41aaeb79a7fbecee27e8b50ac476c990e8241c271ddd4e414
The 4 per-pair input tables, row rules and cohort feature names are copied unchanged from this package.
usesSlow-wave–HFA coupling and HFA power distinguish clinical SOZ depth pairs in 43 patients sci:sha256:9f9a03f5ae81593b8a29528c314db6efb6a4c1b7fb63e4d09e778c7b70330fe7
Source of the ds007118-20 NREM per-pair feature table.
usesAcross 29 SEEG patients, the 30–45 Hz exponent did not separate SOZ channels, while four other features did sci:sha256:115bc8cb0b4c9e536b5f083843a2745508d3c6635a8156b661a9e084f51582a7
Source of the ds003876 NIH multicenter interictal per-pair feature table.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
The archived HUP per-pair feature table used in this analysis is attributed to OpenNeuro ds004100.
usesdoi:10.18112/openneuro.ds003876.v1.0.2
The archived NIH multicenter per-pair feature table used in this analysis is attributed to OpenNeuro ds003876.
usesdoi:10.18112/openneuro.ds005398.v1.1.3
The archived UCLA per-pair feature table used in this analysis is attributed to OpenNeuro ds005398.

Cited by

2 publications

Public catalog entries on Sciscape.

challengesSpike rate and high-frequency activity did not show patient-level complementarity in 39 NREM SEEG patients
Pre-registered test of its non-pre-registered hint that spike rate and HFA fail in different patients; the hint did not replicate in an independent NREM cohort (rho 0.33 vs null 0.31).
supportsAcross 117 SEEG patients in four cohorts, SOZ AUC spread exceeded simulated sampling noise (Q p = 0.040); exploratory true spread was 46%
Both show that per-patient SOZ AUC differences are hard to explain at the patient level; here sampling noise is about half of the AUC variance.