Interictal SEEG features rank seizure-onset-zone contacts in held-out patients and across two clips

Question: can short interictal SEEG recordings find seizure onset zone (SOZ) contacts in a new patient, so that a closed-loop BCI can select its read and stimulation contacts? Data: OpenNeuro ds004100 v1.1.3, 36 SEEG patients, first 150 s of two interictal clips, 3687 bipolar pairs, 392 SOZ pairs. Method: 10 within-patient robust z features (7 log band powers, aperiodic slope, line length, spike-rate proxy) and L2 logistic regression, with leave-one-patient-out validation. Results: median model AUC was 0.786 (35 of 36 patients above 0.5, Wilcoxon p = 2e-10). On the second clip, with the model trained on the first clip of the other patients, the median AUC was 0.811 (35 of 36). Within-shaft AUC was 0.765 (28 of 32), thus the signal is not only a difference between shafts. 28 of 36 patients had one or more SOZ pairs in the top 4 pairs (chance 13.3, p < 1e-5). For resected contacts in Engel 1 patients, the AUC was 0.672 (18 of 20). Single features gave AUC 0.66 to 0.75 with no training. Delta power and spike rate were the best, and the aperiodic slope was flatter in SOZ. The spectral feature maps were very repeatable between clips (Spearman rho 0.96 to 0.99). The spike-rate map was less repeatable (0.77), and the model score had rho 0.84. The AUC did not differ between seizure-free and not seizure-free patients (p = 0.78). Limits: the SOZ labels are clinical and can be partly circular. There are no gray or white matter labels. The data come from one site and are short clips of unknown sleep state. The package contains the code, per-pair features, the S3 manifest with ETags, and the results tables and figures.

SEEG · epilepsy · seizure onset zone · interictal activity · electrophysiology · contact ranking · logistic regression · leave-one-patient-out validation