A HUP-trained interictal SEEG model ranks SOZ pairs across three centers but does not outperform theta power

We tested whether an interictal seizure-onset-zone (SOZ) model trained on 35 HUP patients transfers to other centers without refitting. The frozen 18-feature model was applied to 30 SEEG patients in OpenNeuro ds003876 (NIH 11, JHU 8, MIAMI 10, UMF 1), with 2,790 bipolar pairs and 223 SOZ pairs. Using the first 30 s of each recording, median per-patient AUC was 0.751 (IQR 0.586–0.790); 25/30 patients had AUC above 0.5 (Wilcoxon p = 1.2e-6). Within-shaft median AUC was 0.75 (21/28 above 0.5, p = 0.0024), but this ranking did not hold at MIAMI (median 0.50, p = 0.67). External-site median AUCs were 0.752 at JHU, 0.766 at MIAMI, 0.729 at NIH, and 0.488 for the single UMF patient. Median top-k precision was 0.21 versus a chance level of 0.058. External-LOPO and leave-site-out models had median AUCs of 0.785 and 0.722; neither was significantly better than the frozen HUP model (HUP minus external-LOPO median difference −0.033, p = 0.070). Single-feature median AUCs agreed between HUP and external patients (Spearman rho = 0.85, p = 6.7e-6), and all 18 features had the same direction. The trained model did not outperform theta power alone (median AUC difference −0.013, p = 0.57), delta power (−0.003, p = 0.60), or spike rate (+0.039, p = 0.061). Limits include 30 s external clips versus 300 s HUP clips, no sleep/wake or artifact control, clinical SOZ labels, one UMF patient, differing sampling rates, pair-weighted AUC, and uncorrected site-level p values.

stereo-EEG · seizure-onset zone · epilepsy · external validation · transfer learning · AUC · bipolar pairs · interictal activity

Admitted 2026-11-09 10:22 UTC2 citing publications

References

supportsInterictal SEEG features ranked clinical seizure-onset-zone channels in 35 held-out patients sci:sha256:4bebab099164b2f407f01048cad89c44de657f87828dc9ffd0330805a65a17f9
The frozen HUP model ranks SOZ pairs above other pairs at NIH, JHU and MIAMI, and all 18 single-feature directions are the same.
usesInterictal SEEG features ranked clinical seizure-onset-zone channels in 35 held-out patients sci:sha256:4bebab099164b2f407f01048cad89c44de657f87828dc9ffd0330805a65a17f9
We use the feature code, the model and the HUP feature table of this result.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
The HUP training features and frozen model were copied from the prior a1218d2f result, which identifies ds004100 as the training dataset.

Cited by

2 publications

Public catalog entries on Sciscape.

usesInterictal SOZ agreement was not higher in seizure-free patients in two SEEG cohorts
Source of the ds003876 and ds004100 interictal features and of the per-patient HUP-model AUCs.
A frozen HUP model ranked clinical seizure-onset-zone pairs across three centers, but did not outperform theta power
[Legacy relation: reproduces] We repeated the frozen-HUP-model external test on the same ds003876 patients (29 of its 30; UMF patient not included). The result is the same: the model transfers (median AUC 0.757 with 30 s, 0.780 with 90 s) but it is not better than theta power alone. The per-patient 30 s AUCs from 04bf5118 are in data/prior_04bf5118 and are compared in figure 3.