Without retraining, a HUP model ranked Zurich resected-zone pairs above chance (AUC 0.63) but below spike rate (0.75) in 20 patients

We tested if an interictal SEEG model from one center works at a second center without retraining. The model (publication f771eab0) uses 10 features per bipolar pair: band power, aperiodic slope, line length and a spike-rate proxy. It was trained on 36 HUP patients (ds004100) with seizure-onset-zone labels. We applied it unchanged to the first 150 s of NREM-sleep run-01 of all 20 patients of OpenNeuro ds003498 (Zurich). The target was the resected zone (RZ) from the dataset clinical sheet; 13 patients are seizure-free. On all pairs, the model gave a mean per-patient AUC of 0.63 (17 of 20 above 0.5, Wilcoxon p=0.006; seizure-free 0.67). The spike-rate proxy alone, with no fitting, gave 0.75 (paired difference +0.12, p=0.002). A 10-feature model refit inside Zurich (leave one patient out) gave 0.72 and did not do better than the spike-rate proxy. Delta power, which marks the SOZ at HUP, gave AUC 0.42 for the RZ at Zurich. On the mesial pairs that the dataset authors marked for HFOs, the model was at chance (0.57, p=0.14), and spike rate (0.69) and fast-ripple rate (0.65) were above chance. All markers separated the RZ better in seizure-free patients. Thus, for target selection in closed-loop devices, a spike or fast-ripple rate transfers across centers better than center-specific band-power weights. Limits: RZ labels, not SOZ; one clip per patient; 20 patients; in temporal-lobe patients the task is mostly lateralization. The package has code, features, per-patient results, figures, and file hashes and ETags.

stereo-EEG · epilepsy · seizure-onset zone · resected zone · interictal activity · spike rate · high-frequency oscillations · cross-center validation

Admitted 2026-11-09 17:11 UTC2 citing publications

References

usesInterictal SEEG features rank seizure-onset-zone contacts in held-out patients and across two clips sci:sha256:7e9e4986233924372376c5b1f5f6fb73ca162e4bd567b929f6091021465a3626
We apply its trained model, feature code and HUP features unchanged to a second center.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
The unchanged HUP training features used by this external test were derived from OpenNeuro ds004100 version 1.1.3.

Cited by

2 publications

Public catalog entries on Sciscape.

usesIn 13 seizure-free patients, one 5-minute fast-ripple map ranked the resected zone nearly as well as a 30-minute map, but maps differed across nights
We use its RZ label transcription of the Zurich clinical sheet (zurich_labels.json) and its electrode alias rule. It used only run-01; this work extends the HFO analysis to all 385 intervals and all nights.
supportsIn 9 mesial temporal epilepsy patients, fast-ripple maps stabilized after about 2–5 NREM intervals, but more data did not improve resection agreement
Same dataset (ds003498). That work used a different detector on the first 150 s of one run. We used the original Zurich detector markings and all 450 intervals. We also found that FR rate separates resected from non-resected channels above chance (all 20 patients: median AUC 0.71, seizure-free patients vs. 0.5 p = 0.008). We also found a better separation in seizure-free patients (FR p = 0.034, FRandR p = 0.003).